acceptodds
Under review as a conference paper at ICLR 2027

Web2LongTraj: Scaling Long-Horizon Agent Trajectories from Natural Web Corpora for Mid-Training

25 views

Abstract

Long-horizon agent trajectories are expensive to obtain because most existing pipelines synthesize them in manually engineered executable environments. We present Web2LongTraj, a corpus-first pipeline that mines natural web corpora for real multi-step problem-solving records and compiles them into source-grounded agent trajectories, with actions and outcomes constrained by the source and observations audited for machine-output form. Screening 317 million web records, Web2LongTraj yields 1.15 million audited trajectories. To scale these short episodes into long-horizon supervision, we introduce two operators that grow them along the two axes on which real sessions become long. AnchorStretch extends dependency distances within a trajectory: it first identifies an early premise and a later step that depends on it, then inserts topically similar audited episodes between them. ArtifactStitch composes multi-task sessions by walking a graph that connects trajectories through shared artifacts and compatible working contexts. By recombining the 1.90B-token audited pool, the two operators generate 10.6B long-context tokens, a 5.6-fold expansion without building new task environments. Mid-training Olmo-3-7B and Olmo-3-32B on Web2LongTraj with a standard next-token objective yields consistent gains over equal-budget baselines on tool calling, multi-step interaction, long-horizon memory, and coding benchmarks under both Instruct and Think post-training recipes, while preserving general capability. On Olmo-3-7B, the full Web2LongTraj recipe improves all eight benchmarks by an average of 4.7 points under Instruct SFT and 9.9 points under Think SFT; its five-benchmark general-capability mean also increases by 0.4 and 3.3 points, respectively. These results establish corpus-first trajectory construction as a practical complement to environment-first synthesis for training long-horizon agents.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLMs, mid-training, agentic LLMs, long-context learning, long-horizon agents, synthetic data

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.50 over 4 reviews, above 99% of the others.
Reviewer_gQ1pOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The authors present Web2LongTraj, a pipeline to synthesize long horizon agentic trajectories from web corpora. The automatic data synthesis pipeline created 1.15 million agentic trajectories, and further use AnchorStretch and ArtifactStitch to expand the length and volume of the dataset. The resulting dataset showed gains in a range of agentic and memory benchmarks, compared to the original mid-training data of Olmo-3-7B and Olmo-3-32B.

Critical strengths
  • Fair comparison and solid ablation: The dataset is compared against the original Olmo mid-training data under equal or additive token budget and aligned training parameters. The ablation experiments covered the design choices extensively.
  • Dual focus of agentic and long context: The dataset improves both agentic (tool calling and multi step interaction) and long term memory benchmarks through strategic concatenation of short agentic episodes.
Critical weaknesses
  • The control is non-agentic: the control mixture (Dolma 3 Dolmino and Longmino) contains no tool-use or trajectory data, so Tables 1 measures agentic data versus none in mid-training, instead of an already agentic focused mixture, such as Nemotron-SFT-Agentic-v2 or MidTool. The web-based corpora is therefore claimed to be a complementary source of agentic and long horizon data.
  • Expansion through recombination: the 10.6B long context tokens in the dataset were constructed from 1.90B-token pool through concatenation. AnchorStretch and ArtifactStitch introduced in later training stages contain seen token substrings in earlier stages and training steps.
Questions for authors
  • Could you provide evaluation results of your models on Terminal Bench? It's a long term agentic benchmark that may benefit from diverse knowledge in web corpora.
  • Artifact Stitch is introduced in the last training stage and is probably meant to be of the highest quality and length among the dataset. Why there are no llm-powered filters or refinements on quality or coherence of the multi task trajectory?
4 comments
AuthorsOpenReview ↗

Response to Reviewer gQ1p (1)

We sincerely thank you for your dedicated review and strong endorsement of our paper (4: Clear acceptance), for appreciating our fair comparisons and solid ablations under equal or additive token budgets and aligned training parameters, and for recognizing the dual contribution of our method in simultaneously improving both agentic (tool calling and multi-step interaction) and long-term memory benchmarks.


For Weakness 1

Weakness 1: The control is non-agentic: the control mixture (Dolma 3 Dolmino and Longmino) contains no tool-use or trajectory data, so Tables 1 measures agentic data versus none in mid-training, instead of an already agentic focused mixture, such as Nemotron-SFT-Agentic-v2 or MidTool. The web-based corpora is therefore claimed to be a complementary source of agentic and long horizon data.

Response:

  • Thank you for your suggestion. Regarding Nemotron-SFT-Agentic-v2 and MidTool: the former consists of short-horizon conversational data designed for the SFT stage (data from the same family is already used as a post-training component in our Think SFT mixture), while MidTool, which we discuss and compare in Section 4 (Related Work), has a maximum mid-training context window of only 8K. Neither can directly provide long-horizon mid-training trajectories spanning a 256K context window at a 10B+ token scale, and there is currently no off-the-shelf public long-horizon agent trajectory corpus at a comparable scale and 256K context length for direct comparison.
  • At the same time, our proposed AnchorStretch and ArtifactStitch are source-agnostic expansion operators that can also be directly applied to stretch and stitch trajectories generated from executable environments, which is another important dimension of how our approach complements environment-first synthesis.
  • Furthermore, the MidTool paper also uses an equal-budget Olmo Dolmino mixture (20B tokens) as its mid-training control baseline, which matches our design of using official Olmo 3 data as the control (and our paper validates this at a larger 60B budget).
  • To directly provide a controlled comparison against an agentic corpus executed in real environments, we selected two large public environment-first agent datasets that have zero overlap with our SFT mixtures: TOUCAN (recording multi-turn and parallel tool-execution trajectories across 495 live MCP servers) and Open-SWE-Traces (recording multi-step Docker terminal and code-editing trajectories from OpenHands, SWE-agent, OpenCode, and mini-swe-agent, averaging 31.5K tokens per trajectory). We replaced SeedTraj with an equal 1.90B1.90\text{B}-token volume of TOUCAN trajectories in Stage 2 (8K window), and in Stage 3 we used our AnchorStretch and ArtifactStitch methods to pack a 1:1 token mixture of TOUCAN and Open-SWE-Traces into the 64K / 128K / 256K long-context windows. We then ran experiments using the exact same training (mid-training and post-training) and evaluation setups as Table 1 of the paper. The results are shown below:
Data Recipe (Olmo-3-7B, fixed 60B budget)BFCLIC-Bashτ2\tau^2-benchLongMemEvalAMACLMBPP$+$BCBAvg.
Instruct SFT
Olmo 3 long data23.975.89.023.324.620.345.224.530.8
SeedTraj25.477.610.828.225.222.047.125.432.7
SeedTraj + AnchorStretch29.180.211.029.927.023.750.826.834.8
SeedTraj + ArtifactStitch25.979.512.930.126.822.149.726.334.2
TOUCAN + Open-SWE-Traces29.777.111.929.427.523.147.127.434.2
Web2LongTraj (ours)29.580.612.630.527.823.852.127.535.6
Think SFT
Olmo 3 long data12.660.917.721.712.614.524.119.022.9
SeedTraj23.068.919.726.516.615.833.321.228.1
SeedTraj + AnchorStretch24.370.019.832.620.916.336.021.330.2
SeedTraj + ArtifactStitch23.469.024.026.119.015.635.421.329.2
TOUCAN + Open-SWE-Traces24.876.927.031.420.816.533.321.831.6
Web2LongTraj (ours)25.378.127.233.921.317.236.023.432.8
AuthorsOpenReview ↗

Response to Reviewer gQ1p (2)

As the table shows, TOUCAN + Open-SWE-Traces is a strong public agentic baseline (improving over Olmo 3 long data by +3.4 and +8.7 points under Instruct and Think SFT, respectively, and outperforming single-component ablations in some settings). However, the data synthesized by our full Web2LongTraj pipeline achieves the highest overall average score under both post-training recipes (35.6 vs. 34.2 under Instruct SFT and 32.8 vs. 31.6 under Think SFT, outperforming this baseline on 15 out of 16 benchmark comparisons across the two settings), without requiring manually engineered task environments, MCP servers, or Docker sandboxes. We have added this controlled comparison and analysis to Table 1 and Section 3.2 of the revised paper.


For Weakness 2

Weakness 2: Expansion through recombination: the 10.6B long context tokens in the dataset were constructed from 1.90B-token pool through concatenation. AnchorStretch and ArtifactStitch introduced in later training stages contain seen token substrings in earlier stages and training steps.

Response:

  • What is expanded here is the supervision structure conditioned on long contexts. A short action-observation episode and the exact same episode resumed after many related intervening tasks constitute two different prediction problems: the latter requires the model to resume an earlier task after completing various distracting intermediate tasks without forgetting the information and details of the earlier task. When the same continuation appears after a large number of distracting trajectories rather than inside a short trajectory, the conditional prediction problem that the model must solve is fundamentally different.

The controlled design in Table 1 of the paper directly tests and rules out the hypothesis that "gains come from repeated exposure to the same substrings" while strictly aligning the number of substring repetitions:

Controlled Recipe in Table 1 (Fixed 60B budget, identical 1.90B trajectory pool & Stage 2 overlap)Instruct: 8-Benchmark MeanThink: 8-Benchmark Mean
SeedTraj (Stage 3 uses only official Olmo long documents)32.728.1
↳ w/o dependency split (recombined in equal volume & order as AnchorStretch, without splitting dependencies)32.427.7
SeedTraj + AnchorStretch34.830.2
↳ unrelated sessions (recombined in equal volume from the same pool as ArtifactStitch, randomly concatenated sessions)32.727.1
SeedTraj + ArtifactStitch34.229.2
  1. Dependency-splitting control (w/o dependency split vs. AnchorStretch): The results show that simply packing already-seen trajectories into a long context without stretching dependencies (w/o dependency split) yields no gain in the 8-benchmark average; only when the exact same set of trajectories is placed between the anchor and continuation to form genuine long-range dependencies (AnchorStretch) does the 8-benchmark average improve over SeedTraj by +2.1 / +2.0 points (and outperform w/o dependency split by +2.4 / +2.4 points).
  2. unrelated sessions vs. ArtifactStitch: As shown in the last two rows of the table above, unrelated sessions samples an equal volume of tokens from the same SeedTraj pool and randomly concatenates them into multi-task sessions, yielding no gain over SeedTraj in the 8-benchmark average either, whereas ArtifactStitch with shared-artifact graph walks outperforms unrelated sessions by +1.4 (Instruct) and +2.1 (Think) points. These controls consistently confirm that downstream gains are driven by cross-trajectory dependency and context organization structures rather than repeated substring exposure.
AuthorsOpenReview ↗

Response to Reviewer gQ1p (3)

For Question 1

Question 1: Could you provide evaluation results of your models on Terminal Bench? It's a long term agentic benchmark that may benefit from diverse knowledge in web corpora.

Response:

1. Supplementary Evaluation Results on Terminal-Bench 2.0: Fine-Grained Verification Confirms Significant Cross-Domain Long-Horizon Gains from Web2LongTraj

The latest versions of Terminal-Bench (v4.0/v4.1) specifically removed problems solvable by small and medium-sized models, calibrating single-task difficulty to 2 to 12 hours for a human expert (even 120B-scale models achieve a full-task pass rate of only 0.3% on the official leaderboard). Therefore, to provide a more discriminative comparison for open-weight models at the Olmo-3 scale, we evaluated all five main recipes of Olmo-3-32B (Instruct SFT) from Table 2 of our paper on Terminal-Bench 2.0:

Data Recipe (Olmo-3-32B, 20B Stage 2 + 20B Stage 3)Full-Task Pass@1 (N = 89)Per-Task Partial Score Mean ([0,1] x 100)Official Unit-Test Pass Rate (N = 305)Fraction of Tasks with Effective Progress (>= 1 test passed)
Olmo 3 long data (control baseline)0.0% (0/89)15.919.7% (60/305)39.3% (35/89)
SeedTraj0.0% (0/89)18.721.6% (66/305)47.2% (42/89)
SeedTraj + AnchorStretch1.1% (1/89)21.524.3% (74/305)47.2% (42/89)
SeedTraj + ArtifactStitch0.0% (0/89)21.926.2% (80/305)44.9% (40/89)
Web2LongTraj (Full Method)1.1% (1/89)22.026.6% (81/305)49.4% (44/89)

The experimental results show that Web2LongTraj consistently performs best across all fine-grained execution metrics: Web2LongTraj increases the overall pass rate across all 305 official unit tests from 19.7% to 26.6% and improves the mean per-task partial score from 15.9 to 22.0; on the 55 medium-difficulty tasks that make up the bulk of the benchmark, the mean partial score improves by +9.1 points (14.7 -> 23.8). The number of tasks achieving substantive sub-goal breakthroughs expands from 35/89 to 44/89, and the number of tasks completing more than half of their unit tests increases from 13/89 (14.6%) to 23/89 (25.8%).

2. Why Binary Full-Task Pass@1 Suffers from a Floor Effect, and the Complementary Design of Our Benchmark Suite

The default full-task Pass@1 metric of Terminal-Bench uses a strict "all-or-nothing" criterion (a task receives 1 point only if 100% of its pytest assertions pass, whereas passing 4/5 still receives 0 points), and its official Terminus-2 agent disables native tool-calling interfaces, requiring the model to output strictly escaped JSON keystroke sequences in plain text to drive an interactive tmux terminal. Because our study is designed to strictly isolate the contribution of different mid-training data recipes (fixing a 60B-token mid-training budget at 7B and a 40B-token budget at 32B, followed by general SFT without targeted fine-tuning on any downstream benchmark or specific terminal scaffold), the binary full-task Pass@1 metric is compressed into a 0.0% - 1.1% floor range and cannot sensitively reflect differences between mid-training recipes on its own.

This is precisely why our main experiments systematically combine the following execution-feedback and coding benchmarks:

  • Real command execution and feedback in a Docker Shell (InterCode-Bash): Placing the model in a real Docker container shell to iteratively manipulate the filesystem based on execution feedback under a fair matched budget, Web2LongTraj improves Olmo-3-7B from 75.8 to 80.6 (+4.8) under Instruct SFT and from 60.9 to 78.1 (+17.2) under Think SFT; on Olmo-3-32B, it similarly improves from 81.1 to 84.8 (+3.7) under Instruct SFT and from 70.8 to 80.0 (+9.2) under Think SFT.
  • Multi-step interaction over shared environment state and code transfer (τ2\tau^2-bench, MBPP$+$, and BigCodeBench): On the dynamic environment τ2\tau^2-bench, the 7B model improves by +3.6 / +9.5 and the 32B model improves by +1.9 / +11.4; on the coding benchmarks MBPP$+$ and BigCodeBench, the 7B model improves by +6.9 / +11.9 and +3.0 / +4.4, while the 32B model improves by +3.2 / +6.9 and +0.5 / +1.0 (Instruct / Think SFT), validating the effectiveness of the web-mined trajectories and the two long-horizon expansion operators in Web2LongTraj on real command execution and software engineering tasks.
AuthorsOpenReview ↗

Response to Reviewer gQ1p (4)

For Question 2

Question 2: Artifact Stitch is introduced in the last training stage and is probably meant to be of the highest quality and length among the dataset. Why there are no llm-powered filters or refinements on quality or coherence of the multi task trajectory?

Response:

(1) We would like to clarify that the primary source of ultra-long sequences in our dataset is AnchorStretch (10.00B tokens across 64K-256K windows, with mean dependency distances spanning 52K-227K tokens), which undergoes multi-round independent LLM auditing. In contrast, ArtifactStitch totals 0.11B tokens (0.18% of the 60B total budget), with an average session length of 8.2K tokens, and replaces the official short-document quota in the 256K sub-stage (as noted in Appendix D of our paper, short-context data is mixed in even during the 256K long-context training stage, consistent with the Olmo 3 long-context recipe, to preserve the model's short-context capabilities as the context window grows).

(2) All member trajectories in ArtifactStitch have already passed the three strict LLM gates (Gates 1, 2, and 3, which cumulatively filter out 99.64% of raw web pages). We deliberately avoid generative LLM refinement in order to keep the source records 100.0% byte-for-byte authentic (with 100.0% valid tool calls under the merged tool declaration) and avoid fabricating unverified cross-post intermediate states.

(3) Our independent blind coherence audit of 2,000 randomly sampled ArtifactStitch sessions using gpt-6.1-sol shows that sessions generated by the deterministic artifact-enrichment graph already achieve an 89.8% edge-level shared working-context correctness rate (environment compatibility), with 88.7% having zero cross-task hard conflicts, without any LLM rewriting. Therefore, no additional generative LLM filtering or refinement is needed to achieve both multi-task contextual coherence and byte-for-byte source fidelity.


We sincerely thank you once again for your dedicated review and strong endorsement of our work!

Reviewer_R6wzOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper presents Web2LongTraj, a corpus-first pipeline that mines natural web corpora for multi-step problem-solving records and compiles them into agent trajectories. The paper makes three main contributions: (1) a pipeline that reconstructs and audits agent trajectories from web records without building new task environments; (2) two operators, AnchorStretch and ArtifactStitch, for extending within-task dependency distances and composing related tasks into longer sessions; and (3) controlled mid-training experiments on Olmo-3-7B and Olmo-3-32B under both Instruct and Think SFT, with ablations of the main design choices. The pipeline yields 1.15M audited trajectories and recombines a 1.90B-token pool into 10.6B long-context tokens. On Olmo-3-7B, the authors report average improvements of 4.7 and 9.9 points across eight agent and coding benchmarks under Instruct and Think SFT, respectively, over the equal-budget Olmo 3 long data baseline. The corresponding averages across five general-capability benchmarks also improve by 0.4 and 3.3 points.

Critical strengths
  1. The corpus-first pipeline Web2LongTraj offers a practical way to scale agent training data by reusing problem-solving records already available on the web. It complements environment-first synthesis and reduces reliance on manually engineered environments.
  2. The two operators address complementary aspects of long-context training: recovering distant dependencies within a task and managing context across related tasks. Both introduce useful structure beyond simply concatenating independent texts.
  3. The pipeline includes record screening, constrained trajectory reconstruction, observation auditing, and dependency admission checks. These safeguards address concrete quality issues in converting web records into agent trajectories.
  4. The evaluation is thorough and well controlled, with equal-budget comparisons, two model sizes, two SFT settings, and ablations of key components. Consistent gains across agent and coding benchmarks, alongside preserved general capability, provide strong empirical support.
Critical weaknesses
  1. ArtifactStitch uses shared filenames, environment tags, and semantic similarity to identify related trajectories, but these signals do not establish that file states, tool semantics, and modifications are consistent across the stitched session. The paper should describe session-level consistency checks and report the frequency of state conflicts in an audited sample. A comparison against topically similar task groups without shared state would help separate topical coherence from cross-task state use. In addition, concatenating complete tasks with explicit boundaries captures only part of real long-running work. Reporting task completion rates on interleaved tasks and sessions with ambiguous task boundaries would help assess generalization beyond this structure.
  2. The total cost of data construction remains unclear. Although the pipeline avoids building task environments, scoring, compilation, auditing, and dependency review still require many LLM calls. A breakdown of per-stage call counts, inference tokens, and GPU hours or monetary costs would help quantify the end-to-end construction cost.
  3. The complementary roles of corpus-first construction and environment-first synthesis merit further discussion, particularly in terms of task coverage and access to execution feedback when web records are sparse or lack examples of recovery from failures. Identifying evaluation tools or domains not covered by the training trajectories and reporting performance on those subsets could further clarify the evidence for transfer to unseen environments.
Questions for authors

See Weaknesses above.

4 comments
AuthorsOpenReview ↗

Response to Reviewer R6wz (1)

We sincerely thank you for your dedicated review and strong endorsement of our work (4: Clear acceptance). We greatly appreciate your recognition of our core contributions: scaling agent training data practically by reusing problem-solving records already available on the web without relying on manually engineered environments, complementing environment-first synthesis; introducing useful long-range structure beyond simple concatenation via AnchorStretch (recovering distant within-task dependencies) and ArtifactStitch (managing context across related tasks); addressing concrete quality issues when converting web records into agent trajectories through our quality gates; and providing strong empirical support through well-controlled equal-budget experiments and ablations across two model sizes and two SFT settings while preserving general capabilities.


For Weakness 1

Weakness 1: ArtifactStitch uses shared filenames, environment tags, and semantic similarity to identify related trajectories, but these signals do not establish that file states, tool semantics, and modifications are consistent across the stitched session. The paper should describe session-level consistency checks and report the frequency of state conflicts in an audited sample. A comparison against topically similar task groups without shared state would help separate topical coherence from cross-task state use. In addition, concatenating complete tasks with explicit boundaries captures only part of real long-running work. Reporting task completion rates on interleaved tasks and sessions with ambiguous task boundaries would help assess generalization beyond this structure.

Response:

Session-level consistency checks and frequency of cross-task state conflicts in an audited sample

  • Frequency of state conflicts in the blind audit sample: Our independent blind audit of 2,000 randomly sampled ArtifactStitch sessions using gpt-6.1-sol yields the following results:
    • Edge-level breakdown: 89.8% belong to genuinely compatible shared working contexts; 88.7% have no cross-task hard conflicts.
    • Full-session level: 82.5% of sessions have zero hard conflicts across all adjacent transitions.

Controlled comparison against topically similar task groups without shared state/artifacts

  1. Independent blind audit comparing ArtifactStitch with a strong semantic-neighbor baseline (Semantic-Neighbor-Only): Based on the same Qwen3-Embedding-8B nearest-neighbor graph, we constructed a strong semantic-neighbor baseline (Semantic-Neighbor-Only) that strictly matches ArtifactStitch in scale, session length distribution, and monotonic workflow-stage constraints, without requiring shared filename artifacts. An independent blind audit by gpt-6.1-sol shows that ArtifactStitch achieves an edge-level shared working-context correctness rate (environment compatibility) of 89.8%, outperforming the strong semantic-neighbor baseline Semantic-Neighbor-Only (80.2%) by +9.6 percentage points, and reducing the cross-task hard state contradiction rate from 20.9% to 11.3% (complementing the +1.4 / +2.1 average score improvements of ArtifactStitch over randomly concatenated sessions of the same member trajectories, unrelated sessions, under Instruct / Think SFT in Table 1). Combining shared artifact anchoring, environment compatibility filtering, and enrichment testing allows ArtifactStitch to substantially improve context purity and reduce state conflicts compared to pure semantic neighbors. We have added this blind audit comparison to Appendix C of the revised paper.

Generalization evaluation on interleaved tasks and sessions with ambiguous task boundaries

Real Docker Container Evaluation on Explicit Boundaries, Zero-Marker Ambiguous Boundaries, and Mid-Task Interrupted Interleaved Tasks

To directly test cross-structure generalization in a real executable environment, we constructed three types of multi-task continuous session benchmarks on InterCode-Bash, where tasks share the same persistent Linux container filesystem (each type contains 30 3-task continuous sessions = 90 real Docker tasks per condition, maintaining a single continuous multi-turn conversation context throughout each session, for 270 executed container tasks per model):

  • Explicit Boundary: Consecutive tasks include explicit task-switch delimiter prompts between them;
  • Ambiguous Boundary (zero boundary markers): Immediately after the final tool observation of the previous task returns, the next user task request enters the same conversation directly, with all boundary cues completely removed;
AuthorsOpenReview ↗

Response to Reviewer R6wz (2)

  • Interleaved Tasks (mid-task interruption and recovery): Task A performs its initial round of exploration and is then interrupted by the full execution of Task B; subsequently, within the same live container state, the agent must resume and complete Task A from a brief coreference prompt before proceeding to Task C.

The table below reports the real container execution performance (InterCode-Bash Mean Reward and full-score Solve Rate, n = 90 tasks per condition) for the core fixed 60B-budget checkpoints from Table 1 across the three session structures:

60B Matched-Budget ModelExplicit Boundary Mean Reward / Solve RateAmbiguous Boundary (Zero Markers) Mean Reward / Solve RateInterleaved Tasks (Mid-Task Interruption) Mean Reward / Solve RateMean Reward on Resumed Interrupted Tasks (n = 30)
Instruct SFT
Olmo 3 long data77.2 / 30.0%77.9 / 32.2%78.6 / 33.3%80.3
SeedTraj77.5 / 36.7%79.0 / 36.7%79.3 / 34.4%77.2
SeedTraj + ArtifactStitch78.8 / 37.8%78.0 / 31.1%77.5 / 30.0%78.2
↳ Unrelated sessions77.9 / 36.7%76.2 / 28.9%77.0 / 27.8%77.4
Web2LongTraj (Full Recipe)81.4 / 42.2%81.1 / 37.8%80.6 / 33.3%80.4

As the results show, ArtifactStitch improves over Unrelated sessions (which concatenates the exact same sub-trajectories) on both unmarked-boundary sessions and mid-task interrupted interleaved sessions; further, backed by the complementary design of AnchorStretch (interruption and recovery across intervening tasks) and ArtifactStitch (multi-task switching in a shared context), Web2LongTraj generalizes to both zero-marker ambiguous-boundary sessions and interleaved tasks.


For Weakness 2

Weakness 2: The total cost of data construction remains unclear. Although the pipeline avoids building task environments, scoring, compilation, auditing, and dependency review still require many LLM calls. A breakdown of per-stage call counts, inference tokens, and GPU hours or monetary costs would help quantify the end-to-end construction cost.

Response:

The table below summarizes the data volume processed, estimated request counts, total inference input/output tokens, and H20 GPU hours across all six stages of the Web2LongTraj pipeline:

Pipeline StageDeployed ModelProcessed Data Scale & Retained OutputEstimated RequestsEstimated Input / Output TokensEstimated H20 GPU Hours (Share)
Gate 1: Raw web readiness scoringQwen3.6-35B-A3B317.4M raw records →\to 6.68M seeds (2.1%)~320.0M~800.0B / ~58.0B~8,500 (30.2%)
Gate 2: Web record compilation into short trajectoriesGLM-5.2-FP86.68M seeds →\to 1.89M multi-step short trajectories (28.3%)~7.3M~20.5B / ~9.9B~15,000 (53.3%)
Gate 3: Trajectory observation-form auditingGLM-5.2-FP81.89M trajectories (6.67M obs) →\to 1.15M SeedTraj (1.90B tokens)~2.0M~2.4B / ~0.06B~1,200 (4.3%)
AnchorStretch: Splitting, probes, independent review & Δ\Delta testGLM-5.2-FP81.15M short trajectory pool →\to admitted long-range dependency pairs & probes~2.5M~5.9B / ~0.60B~3,400 (12.1%)
Semantic embedding & exact nearest-neighbor indexingQwen3-Embedding-8B + Faiss-GPUTask-signature encoding & full-pool exact KNN for 1.15M SeedTraj trajectories~1.2M~0.5B / -~20 (0.1%)
Long-context assembly & ArtifactStitch graph walksNo LLM (pure CPU deterministic execution)Produces 113.0K long-range sequences + 13.2K multi-task sessions (10.6B tokens total)0- / -0 (~50 CPU core-hours)
End-to-End Total (Approximate)-1.15M audited short trajectories (1.90B) + 10.6B long-context tokens~333.0M~829.3B / ~68.6B~28,120 H20 hours

Note on the approximate values reported in the table: In actual large-scale production, inference was executed asynchronously through a shared multi-user elastic inference service pool (where service instances dynamically scaled with cluster load, were shared across services, and ran on heterogeneous underlying inference GPU pools), and actual prefix-caching hit rates varied dynamically with load-balancing routing, making it difficult to collect exact statistics.

Zero generative cost for long-context expansion itself: The long-context assembly of AnchorStretch and the ArtifactStitch method involve only deterministic retrieval, graph walks, and byte-level concatenation. This completely avoids the high compute cost and latency of running multi-turn long-horizon teacher rollouts over 64K to 256K context windows.

AuthorsOpenReview ↗

Response to Reviewer R6wz (3)

For Weakness 3

Weakness 3: The complementary roles of corpus-first construction and environment-first synthesis merit further discussion, particularly in terms of task coverage and access to execution feedback when web records are sparse or lack examples of recovery from failures. Identifying evaluation tools or domains not covered by the training trajectories and reporting performance on those subsets could further clarify the evidence for transfer to unseen environments.

Response:

1. Complementary Roles of Corpus-First Construction and Environment-First Synthesis, and Quantification of Failure Recovery

The core distinction between the two paradigms lies in how execution feedback is obtained and their coverage distributions, making them highly complementary across training stages. In technical ecosystems with rich public troubleshooting records, Web2LongTraj directly extracts the real "action -> observation -> repair" feedback loops already preserved in source records (Gates 1 to 3 specifically discard operational advice that lacks execution results), covering long-tail software stacks at low cost without building per-post sandboxes. This establishes a clear division of labor: using corpus-first data during mid-training to inject broad priors for tool interaction, error recovery, and long-horizon state tracking, thereby allowing costly environment engineering to focus on proprietary interfaces and post-training policy alignment. Our Think SFT experiments (Web2LongTraj mid-training followed by SFT on mixtures containing environment-synthesized agent datasets such as Nemotron-SFT-Agentic, OpenThoughts-Agent, and xLAM) directly instantiate this combination of the two paradigms and achieve the largest overall gains (+9.9 at 7B and +5.8 at 32B). Moreover, as source-agnostic operators, AnchorStretch and ArtifactStitch can be applied directly to extend environment-synthesized trajectories as well.

Failure and error recovery is not a weakness of web corpora, but a core component of Web2LongTraj. Community technical posts exist precisely because execution encountered errors or unexpected outcomes, and Gate 1 reserves its highest readiness scores specifically for records that document observation logs or error-repair processes. An exhaustive census across all 1,154,483 trajectories in SeedTraj shows that 34.6% contain intermediate failure or bug-reproduction observations, followed by identifying the root cause in <think> reasoning and issuing corrective tool calls. Specifically, 21.7% explicitly contain literal machine error outputs in <tool_response> (exception tracebacks, compilation errors, HTTP 4xx/5xx, SQL errors, or non-zero exit codes), while another 12.9% reproduce and repair silent semantic errors (such as incorrect numerical values, empty query results, or formatting mismatches).

2. Transfer Performance on Tool Families and Unseen Environment Subsets Not Covered by Web2LongTraj

Here we strictly define "Uncovered" as not covered by the newly constructed Web2LongTraj mid-training corpus. We first normalize API functionalities and tool families semantically, and then combine environment domains with the maximum Qwen3-Embedding-8B cosine similarity of each evaluation item relative to the SeedTraj corpus to partition the evaluation subsets, reporting subset scores and gains under the exact same setup as Table 1:

  • BFCL-V4: We group POSIX filesystem/OS, math/scientific computing, SQL, and web search into Web2LongTraj Covered Tool Families (33.5% of official total weight); we group vehicle control, stock trading, travel booking, ticketing/social media, consumer lifestyle APIs, and the three persistent memory backends (memory_kv/vector/rec_sum) into Web2LongTraj Uncovered Tool Families (66.5% of official total weight). Taking the strict intersection "Uncovered tool family ∩\cap normalized function name absent from SeedTraj ∩\cap maximum cosine similarity to SeedTraj below the full-set median of 0.688" forms a Triple-Strict Unseen Subset (37.0% of official total weight).
  • τ2\tau^2-bench: Across all tools in the airline, retail, and telecom domains, apart from the basic arithmetic helper calculate, 100% of the proprietary business-state query and modification APIs never appear in Web2LongTraj; among them, the telecom domain (such as line suspension/resumption, data top-up, and roaming toggles) has zero tool overlap and accounts for 41.0% of all tasks.
  • AMA-Bench: Partitioned into Web2LongTraj Covered Technical Domains (SWE-bench, Spider2 SQL, and GAIA, accounting for 56.2%) and Web2LongTraj Zero-Coverage Heterogeneous Interactive Environments (WebArena web GUI clicking, 5 symbolic grid games, and ALFWorld embodied household rooms, accounting for 43.8%).
AuthorsOpenReview ↗

Response to Reviewer R6wz (4)

Benchmark & Coverage Partition (Olmo-3-7B, fixed 60B budget, aligned with Table 1)Instruct: Olmo 3 long →\to Web2LongTrajInstruct Gain Δ\DeltaThink: Olmo 3 long →\to Web2LongTrajThink Gain Δ\Delta
BFCL-V4 Full Set (100% weight, Table 1)23.9 →\to 29.5+5.612.6 →\to 25.3+12.7
↳ Web2LongTraj Covered Tool Families (33.5% weight)18.0 →\to 21.2+3.211.9 →\to 21.5+9.6
↳ Web2LongTraj Uncovered Tool Families (66.5% weight)26.9 →\to 33.7+6.812.9 →\to 27.2+14.3
↳ Triple-Strict Unseen Intersection Subset (37.0% weight)24.6 →\to 30.5+5.99.5 →\to 21.7+12.2
τ2\tau^2-bench Full Set (100% proprietary business APIs unseen, Table 1)9.0 →\to 12.6+3.617.7 →\to 27.2+9.5
↳ telecom Domain (41.0% share, zero tool overlap)7.7 →\to 10.3+2.611.2 →\to 31.1+19.9
AMA-Bench Full Set (100% tasks, Table 1)24.6 →\to 27.8+3.212.6 →\to 21.3+8.7
↳ Web2LongTraj Covered Technical Domains (56.2%)25.9 →\to 28.9+3.014.7 →\to 25.1+10.4
↳ Web2LongTraj Uncovered Heterogeneous Environments (43.8%)23.1 →\to 26.5+3.49.8 →\to 16.4+6.6

The results demonstrate that Web2LongTraj exhibits strong generalization and transfer to tool families and environments not covered by the training trajectories, rather than merely imitating seen tools and environments: it achieves substantial gains across all uncovered subsets of all three benchmarks under both SFT settings. In fact, on BFCL-V4, the gains on uncovered tool families (+6.8 / +14.3) exceed those on covered tool families (+3.2 / +9.6), and the zero-tool-overlap telecom dual-control environment improves by +19.9 points under Think SFT (at the 32B scale, τ2\tau^2-bench also improves by +1.9 / +11.4). This provides strong evidence that Web2LongTraj equips the model with transferable foundational capabilities for interpreting newly declared tool schemas, using environmental feedback to correct errors, and tracking state over long contexts, rather than memorizing specific tools in the training corpus.


We sincerely thank you once again for your dedicated review and strong support of our work!

Reviewer_1gzoOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Web2LongTraj, a corpus-first pipeline that constructs long-horizon agent trajectories from in-the-wild web corpora. The authors first filter 317 million web records to select those containing actions and outcomes. The selected records are then converted into ReAct-formatted agent trajectories. Based on these short trajectories, the authors expand them into long-horizon and multi-task datasets, respectively. The authors demonstrate the effectiveness of this data curation pipeline through mid-training experiments, showing consistently better results compared to Olmo 3 long data.

Critical strengths
  1. This paper has clear motivation to this data construction pipeline and presents a practical and scalable data curation solution with identification of the cost for building data from executable environments. The paper also has relatively comprehensive experimental evaluation and well-designed ablations, including model sizes, training settings, multiple evaluation benchmarks and general capability benchmarks.
Critical weaknesses
  1. My first concern is about the reliability and faithfulness of the reconstructed agent trajectories. Since the pipeline does not use any executable environments, it heavily relies on LLMs to reconstruct the agent trajectories from natural language. Although there are screenings, the audit proposed by the author mainly evaluates the trajectory formats rather than outcome correctness. Regarding reconstruction, the authors do not demonstrate if this LLM reconstruction can faithfully convert the interactions. A human evaluation comparing sampled trajectories against their sources would help validate the quality of the constructed dataset.
  2. A minor concern is the limited justification for using the constructed trajectories specifically during mid-training rather than SFT. Although the authors demonstrate improvements through mid-training, a comparison with SFT-based training would help clarify the benefits of this design choice.
Questions for authors
  1. Given the large-scale web corpus used for trajectory construction, did the authors conduct any de-contamination checks to ensure that the training data does not overlap with downstream evaluation benchmarks?
4 comments
AuthorsOpenReview ↗

Response to Reviewer 1gzo (1)

We sincerely thank you for your careful review and positive feedback. We are glad that you find our motivation clear, our pipeline a "practical and scalable data curation solution," our experimental evaluation comprehensive, and our ablations across model sizes, training settings, and benchmarks well-designed.


For Weakness 1

Weakness 1: My first concern is about the reliability and faithfulness of the reconstructed agent trajectories. Since the pipeline does not use any executable environments, it heavily relies on LLMs to reconstruct the agent trajectories from natural language. Although there are screenings, the audit proposed by the author mainly evaluates the trajectory formats rather than outcome correctness. Regarding reconstruction, the authors do not demonstrate if this LLM reconstruction can faithfully convert the interactions. A human evaluation comparing sampled trajectories against their sources would help validate the quality of the constructed dataset.

Response:

Thank you for raising this important concern. We first clarify why Web2LongTraj performs a source-constrained representation conversion of recorded interactions rather than open-ended LLM simulation or free reconstruction, and then report a detailed item-by-item evaluation comparing sampled trajectories directly against their original web pages.

1. Three-Stage Source-Fact Constraint Mechanism Beyond Surface Format Checks

Although our pipeline does not instantiate isolated execution containers for each post, it does not invent tasks out of thin air from arbitrary natural language narratives or ask an LLM to guess unrecorded execution outcomes. Data quality is jointly guaranteed by three gates with complementary roles:

  • Gate 1 (Source Record Observability Screening): Out of 317.4M quality-prefiltered web documents, 97.9% (about 310.7M / 317.4M) are directly eliminated because they lack explicit action procedures and observed outcomes (ARS < 4), retaining only 2.1% (6.68M / 317.4M seeds) that already explicitly document multi-step actions and results to enter the compilation stage.
  • Gate 2 (Source-Constrained Compilation and "Rejection Over Patching"): The compiler (GLM-5.2-FP8) is strictly required to preserve the actual commands, arguments, numerical values, error types, and final solutions reported in the source post, drop steps with indeterminate outcomes, and directly reject rather than patch entire seeds that contain only operational advice without execution evidence or that would require fabricating core facts to form a trajectory. This stage eliminates 71.7% (about 4.79M / 6.68M) of the seeds, retaining only 28.3% (1.89M / 6.68M trajectories).
  • Gate 3 (Observation Semantic-Role Auditing): When Gate 3 inspects <tool_response>, it does not merely check JSON or tag syntax; rather, it rejects post-hoc human author summaries disguised as tool returns based on their semantic role (for example, a syntactically valid return stating "this change fixed the problem" is rejected; this flags 22.26% or about 1.48M / 6.67M observations and rejects 39.02% or 738,859 / 1,893,342 compiled trajectories). Combining all three gates, only 0.36% (1.15M / 317.4M trajectories, containing 3,821,539 observations) of the initial 317.4M web records enter SeedTraj.
AuthorsOpenReview ↗

Response to Reviewer 1gzo (2)

2. Stratified Faithfulness Evaluation Against Original Web Pages

Following your suggestion, we conducted an item-by-item side-by-side comparison of sampled SeedTraj trajectories against their original web texts. To ensure that different source coverage intervals are thoroughly audited, we partitioned all 1,154,483 trajectories into four buckets according to the character-level 20-gram coverage of their observations relative to the source text ([0.0, 0.1): 30.1%; [0.1, 0.4): 24.0%; [0.4, 0.7): 24.0%; [0.7, 1.0]: 21.9%), uniformly sampled 750 trajectories at random from each bucket (totaling N = 3,000 trajectories and 9,970 steps), paired 100.0% of them back to their exact original source posts, and weighted the results by each bucket's exact proportion in the data pool to estimate full-pool metrics.

Each (original web page, compiled trajectory) pair was verified item by item across three granularities (trajectory level, action level, and observation level) by an independent evaluator model from a different model family than the GLM-5.2 compiler (gpt-6.1-sol) under a blind setting with bucket information hidden.

We validated the reliability of this evaluation protocol through two independent calibrations:

(i) On 400 sentinel samples blindly mixed into the review queue (300 with planted factual contradictions in outcome polarity, error type, or output numerical values, plus 100 unmodified clean controls), the target-step contradiction detection rate reached 96.7%, with a 0.0% false positive rate on the target steps of clean controls;

(ii) On a stratified subset of 600 trajectories (2,000 steps in total, 150 trajectories per bucket), we conducted item-by-item human verification against the original web pages; the four-class agreement rate between the primary evaluator and human verification reached 98.5%, and the binary agreement rate on source grounding reached 99.0%.

Source Faithfulness Metric (SeedTraj pool-weighted; stratified sample N = 3,000 trajectories, 9,970 steps)Pool-Weighted Proportion
Trajectory-Level Interaction Faithfulness
Task description (User query) faithfully reflects the core problem in the source post100.0%
Causal step progression order is faithful to the source record94.6%
No ungrounded or fabricated substantive action steps introduced93.3%
No key steps omitted from the solution in the source post96.5%
Trajectory reproduces the actual final resolution/conclusion in the source post92.5%
Action-Level & Observation-Level Factual Grounding
Actions with source-grounded commands (including 78.2% direct/paraphrased + 21.0% file-wrapped source snippets)99.2%
Actions with both commands and arguments grounded in the source97.0%
Steps whose success/failure outcome polarity matches the source record95.1%
Total observations grounded in or plausibly consistent with the source (32.8% verbatim + 38.3% uniquely reconstructable + 22.3% plausible runtime surface reconstruction)93.4%
Observations containing unsupported or source-contradictory facts6.6%

3. Sandbox Execution Verification on Self-Contained Actions and Controlled Downstream Training Validation

As a supplementary execution verification, we identified a replayable subset of 1,000 trajectories and replayed them sequentially in completely network-isolated Docker containers.

Execution Verification Metric (Replayable subset: N=1,000N = 1,000 trajectories, 3,0803,080 executed command steps)Agreement Rate with Actual Execution
Command agreement (matches record and executes validly)97.5%
Argument agreement (matches record and executes validly)96.8%
Outcome agreement (execution outcome polarity matches actual execution)92.4%

In this replayable subset, 97.5% of commands execute validly as recorded, 96.8% of arguments agree with actual execution requirements, and 92.4% of execution outcome polarities agree with actual execution.

Finally, these direct source-page comparisons and sandbox replay results align closely with our controlled ablation experiments in Table 1 under strictly fixed source documents and training token budgets. Holding the source records and token budget fixed, Table 1 compares raw web text, trajectories without observation auditing, and SeedTraj. SeedTraj consistently outperforms both baselines across all eight benchmarks under Instruct SFT (8-benchmark average of 32.7, outperforming 31.1 for raw web text and 30.8 for unaudited trajectories) and Think SFT (8-benchmark average of 28.1, outperforming 26.4 for raw web text and 25.8 for unaudited trajectories). This confirms that our data processing effectively improves the supervision signal available for training.

AuthorsOpenReview ↗

Response to Reviewer 1gzo (3)

For Weakness 2

Weakness 2: A minor concern is the limited justification for using the constructed trajectories specifically during mid-training rather than SFT. Although the authors demonstrate improvements through mid-training, a comparison with SFT-based training would help clarify the benefits of this design choice.

Response:

  • We prioritize positioning the data synthesized by Web2LongTraj in the mid-training stage for the following reasons:
    • Scale and base model reuse: Mid-training is the stage capable of absorbing data of this scale and context length. Because web corpora are massive, our web-corpus synthesis pipeline generated 10B+ tokens in our experiments alone. This volume is too large for standard SFT; if added directly to SFT, it would overwhelm other carefully balanced SFT mixtures.
    • Foundation model training workflow: Post-training typically uses a smaller amount of highest-quality trajectory data, whereas web corpora can synthesize a very large volume of trajectories, making them well suited for the mid-training stage as a warm-up before post-training. Web2LongTraj provides broad coverage and is best placed in a stage aimed at capability exposure and mixed with general pre-training data, leaving SFT for fine-grained behavioral alignment. Injecting large-scale agent behavioral data during mid-training is also common practice in the community, and prior agentic mid-training works (such as GLM-4.5, daVinci-Dev, and MidTool) are motivated by the same considerations.

To directly quantify the difference between placing Web2LongTraj in mid-training versus SFT, we conducted a controlled comparison where we simply moved the Web2LongTraj data from the mid-training stage to the SFT stage while keeping the remaining mid-training budget and the original SFT budget unchanged. The results are shown below:

Data Recipe (Olmo-3-7B, fixed 60B budget)BFCLIC-Bashτ2\tau^2-benchLongMemEvalAMACLMBPP$+$BCBAvg.
Instruct SFT
Olmo 3 long data23.975.89.023.324.620.345.224.530.8
SeedTraj25.477.610.828.225.222.047.125.432.7
SeedTraj + AnchorStretch29.180.211.029.927.023.750.826.834.8
SeedTraj + ArtifactStitch25.979.512.930.126.822.149.726.334.2
Web2LongTraj (SFT)29.180.312.930.427.023.951.627.035.3
Web2LongTraj (Mid-Training)29.580.612.630.527.823.852.127.535.6
Think SFT
Olmo 3 long data12.660.917.721.712.614.524.119.022.9
SeedTraj23.068.919.726.516.615.833.321.228.1
SeedTraj + AnchorStretch24.370.019.832.620.916.336.021.330.2
SeedTraj + ArtifactStitch23.469.024.026.119.015.635.421.329.2
Web2LongTraj (SFT)24.476.727.334.021.016.935.723.032.4
Web2LongTraj (Mid-Training)25.378.127.233.921.317.236.023.432.8

As the table shows, training on our data in SFT performs comparably to mid-training, while placing the data in the mid-training stage yields slightly better overall performance than placing it in SFT (an 8-benchmark average of 35.6 vs. 35.3 under Instruct SFT, and 32.8 vs. 32.4 under Think SFT, outperforming SFT on 6 out of 8 benchmarks in both settings). This is because full-sequence next-token prediction during mid-training exposes the model to user instructions and environmental observations as well: the model not only learns how to invoke tools, but also models environmental feedback and state transitions as a world model, similar to standard pre-training. We have added this comparison and analysis to Appendix I and Table 10 (referenced in Section 3.1) of the revised paper.

AuthorsOpenReview ↗

Response to Reviewer 1gzo (4)

For Question 1

Question 1: Given the large-scale web corpus used for trajectory construction, did the authors conduct any de-contamination checks to ensure that the training data does not overlap with downstream evaluation benchmarks?

Response:

We compared (a) the 1.90B-token raw web source records, (b) the complete Web2LongTraj corpus (1.15M SeedTraj trajectories, all 18,392 unique AnchorStretch planted probes, and all 13,164 ArtifactStitch sessions), and (c) 100% of the examples in both original open-source SFT mixtures against all 13 evaluation benchmarks, using the following methods:

  1. Normalized n-gram overlap: Exact 13-gram matching on full training documents (excluding generic template phrases that appear in ≥5\ge 5 evaluation problems or >10> 10 corpus documents; full-sequence exact matching for short prompts of 8 to 12 words) as well as 50-token contiguous window matching.
  2. MinHash-LSH near-duplicate detection: 128-permutation MinHash-LSH on normalized 5-gram shingles with exact Jaccard verification at J≥0.80J \ge 0.80.
  3. Semantic nearest-neighbor retrieval: Qwen3-Embedding-8B semantic nearest-neighbor search over the full trajectory vector index and planted probes, with item-by-item manual inspection of all candidate pairs with cosine similarity ≥0.85\ge 0.85.

We manually reviewed all candidate pairs with cosine similarity ≥0.85\ge 0.85 together with all nn-gram and MinHash candidates against the specific constraints, problem statements, and reference solutions of each benchmark to verify whether they constituted genuine overlap.

The results are as follows:

Across all benchmark evaluation items, the data synthesized by our Web2LongTraj pipeline has 0 verified task-level overlaps across all eight agent/coding benchmarks and all five general-capability benchmarks (the number of items satisfying MinHash Jaccard ≥0.80\ge 0.80 on every benchmark is 0; manual review of all 309 embedding nearest-neighbor pairs with cosine similarity above 0.85 confirmed that although their embedding similarity is relatively high, they are actually independent tasks with different input arguments, tool schemas, and execution outcomes, while surface n-gram collisions originate from public-domain literary test strings or standard library/array constants and do not constitute data leakage).

In the raw web source records, only 1 BFCL prompt and 2 MMLU question stems overlapped, and all three records were rejected by Gates 1 to 3 due to the lack of actual executed action-observation loops and never entered the training data.

Meanwhile, the exact same detector accurately captured a small number of overlaps within the original open-source SFT mixtures (7/4,696 generic short queries on BFCL; on MBPP+, 46/378 derived prompts originating from the original MBPP training split in the released Dolci-Instruct-SFT and 4/378 in Think SFT; and 0 on the remaining six core benchmarks), which validates the detection sensitivity. We already filtered out these overlapping examples in the SFT recipes actually used for training. We have added this decontamination audit procedure and its results to Appendix E.1 (and referenced it in Section 3.1) of the revised paper.


We sincerely thank you once again for your dedicated review and valuable comments! If our responses and additional experiments have addressed your concerns, we would be grateful if you would consider raising your score.

Reviewer_DvvkOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Web2LongTraj constructs agent-style mid-training data from natural web records. SeedTraj compiles action-oriented records into ReAct trajectories and filters observations for machine-output form. AnchorStretch increases the distance between existing dependencies by inserting other trajectories, while ArtifactStitch joins related tasks into longer sessions. Fixed-budget mid-training experiments, followed by consistent SFT within each condition, report gains across interaction, memory, and coding benchmarks at 7B and 32B scales. The paper also studies dependency structure, auditing, and task composition through ablations.

Critical strengths
  1. The matched-budget experiments support the practical value of the data recipe. At 7B with a fixed 60B-token budget, the full recipe improves all eight reported benchmarks within both SFT conditions, with mean gains of 4.7 and 9.9 points; the 32B results also improve on their respective controls (Tables 1–2). This breadth is stronger evidence than merely showing that the generated text resembles agent interaction. The aggregate remains a descriptive average of different benchmark metrics.

  2. The dependency ablation is particularly informative. Holding member trajectories and their order fixed while changing whether a dependency is separated yields LongMemEval scores of 29.9 versus 20.1 and 32.6 versus 24.5 (Table 1). This helps distinguish the benefit of dependency structure from simply supplying longer text. Same-source and same-budget comparisons also make the data contribution more interpretable.

Critical weaknesses
  1. Observation-form auditing does not establish source fidelity or execution validity. Gate3 inspects whether an observation looks like machine output. Source n-gram coverage is not an admission threshold, and compilation may reconstruct output when the source describes a result rather than providing a log (Appendices A and I.2). The stated prohibition on new facts is appropriate, but its reliability at scale is not independently measured. ArtifactStitch's semantic enrichment likewise does not establish compatible file contents, versions, paths, or preconditions. This limits the interpretation of the data as realistic executable experience, even if it is effective synthetic training material.

  2. Uncertainty and provenance checks are insufficient for the finer conclusions. Temperature-zero evaluation does not eliminate training variation, finite-test uncertainty, or simulator/judge error. Small component differences lack multi-seed or paired uncertainty estimates, and the claimed agreement with the official proprietary judge needs numerical calibration details (Appendix E). I also did not identify a systematic contamination audit covering source records, the evaluation benchmarks, and SFT data. This is an unresolved validation gap, not evidence that contamination occurred.

Questions for authors
  1. Can an independently annotated sample quantify verbatim observations, uniquely reconstructable outputs, underdetermined but plausible reconstructions, and unsupported facts? For any replayable subset, what fractions of commands, arguments, and outcomes agree with actual execution?
  2. What are the source collection dates and deduplication procedures against benchmark questions, code solutions, and SFT data? Please clarify the evidence for excluding near-duplicate contamination.
  3. What is the absolute correctness rate of ArtifactStitch edges and the rate of cross-task state contradictions? How does it compare with a strong semantic-neighbor-only composition baseline? The reported enrichment factor should not be interpreted as edge correctness.
  4. Please provide paired uncertainty estimates for key ablations, judge comparison sample sizes and disagreements, and the existing-dependency/planted-probe mixture and trajectory reuse distribution. Evaluation on tasks requiring repeated state updates and recovery would be useful future validation; the current results should be framed as downstream benchmark gains rather than proof of prolonged autonomous work.
6 comments
AuthorsOpenReview ↗

Response to Reviewer Dvvk (1)

We sincerely thank you for your thorough review and detailed feedback. We are glad that you find our matched-budget experiments across eight benchmarks and two model scales supportive of the practical value of our data recipe, and our dependency ablation informative in distinguishing the benefit of dependency structure from simply supplying longer text.


For Weakness 1 (Part I) & Question 1

Weakness 1 (Part I): Observation-form auditing does not establish source fidelity or execution validity. Gate3 inspects whether an observation looks like machine output. Source n-gram coverage is not an admission threshold, and compilation may reconstruct output when the source describes a result rather than providing a log (Appendices A and I.2). The stated prohibition on new facts is appropriate, but its reliability at scale is not independently measured.

Question 1: Can an independently annotated sample quantify verbatim observations, uniquely reconstructable outputs, underdetermined but plausible reconstructions, and unsupported facts? For any replayable subset, what fractions of commands, arguments, and outcomes agree with actual execution?

Response:

Thank you for recognizing that the trajectories are effective synthetic training material. We would like to clarify the following points:

  • Our quality control comes from the complete three-stage pipeline rather than relying solely on Gate 3 to check whether an output "looks like machine text." Source fidelity is strictly enforced by Gate 1 and Gate 2 (the first two gates together filter out 99.4%99.4\% of raw web records, and all three gates cumulatively filter out 99.64%99.64\%, retaining only 0.36%0.36\%), whereas the specific role of Gate 3 is to remove human narrative formatting mixed into tool responses. In our pipeline, Gate 3 does not perform fact verification in isolation; instead, it builds upon the source-fact constraints of Gate 1 and Gate 2. Because Gate 3 inspects only the <tool_response> text, it audits the surface output form of observations rather than factual fidelity or execution validity. In Web2LongTraj, factual provenance constraints and surface-form auditing are handled by separate stages of the pipeline. Gate 1 filters 317.4M pre-filtered web records down to 2.1% (6.68M / 317.4M seeds satisfying ARS >= 4, eliminating 97.9% or about 310.7M / 317.4M records), ensuring that records entering subsequent stages already document concrete multi-step actions and observed outcomes. Gate 2 (GLM-5.2-FP8) compiles these records under strict source constraints, requiring the compiler to preserve commands, arguments, values, and results reported in the source post, drop steps with indeterminate outcomes, and reject entire records that would require fabricating core facts to form a trajectory rather than forcibly patching them (eliminating 71.7% or about 4.79M / 6.68M seeds, and retaining only 28.3% or 1.89M / 6.68M trajectories). Subsequently, Gate 3 specifically removes an independent failure mode, namely post-hoc human narrative summaries mixed into tool returns (flagging 22.26% or about 1.48M / 6.67M narrative observations across 6.67M observations; because a single narrative observation vetoes the entire trajectory, this rejects 39.02% or 738,859 / 1,893,342 compiled trajectories, ultimately retaining 60.98% or 1,154,483 / 1,893,342 trajectories with 3,821,539 observations/steps in SeedTraj).
  • Regarding the concern that "the stated prohibition on new facts is appropriate, but its reliability at scale is not independently measured," we conducted controlled comparisons in Table 1 and provide a direct, independent paired audit below. Table 1 compares raw web text, trajectories without observation auditing, and SeedTraj while holding the source records and token budget fixed. SeedTraj consistently outperforms both baselines across all eight benchmarks under Instruct SFT (8-benchmark average of 32.7, outperforming 31.1 for raw web text and 30.8 for unaudited trajectories) and Think SFT (8-benchmark average of 28.1, outperforming 26.4 for raw web text and 25.8 for unaudited trajectories). This demonstrates that our data processing effectively improves the supervision signal available for training.
AuthorsOpenReview ↗

Response to Reviewer Dvvk (2)

  • Furthermore, we record character-level n-gram source coverage as per-trajectory metadata rather than setting it as an admission threshold because copying a human comment verbatim (such as "it worked") yields 100% (1.0) coverage without being valid machine output, whereas converting an outcome stated in natural language in the source into a standardized tool output lowers literal surface overlap without changing the factual content. The exact distribution across the four trajectory coverage buckets is: [0.0, 0.1) accounts for 30.1% of trajectories (346,967 / 1,154,483 trajectories; containing 28.8% or 1,101,996 / 3,821,539 observations), [0.1, 0.4) accounts for 24.0% (277,226 / 1,154,483 trajectories; containing 26.8% or 1,025,159 / 3,821,539 observations), [0.4, 0.7) accounts for 24.0% (277,474 / 1,154,483 trajectories; containing 23.3% or 888,600 / 3,821,539 observations), and [0.7, 1.0] accounts for 21.9% (252,816 / 1,154,483 trajectories; containing 21.1% or 805,784 / 3,821,539 observations).

To directly answer the two sub-questions of Question 1, we conducted two independent evaluations: a stratified source-paired audit (Test 1) and an isolated Docker sandbox execution replay on a self-contained subset (Test 2).

Regarding "Can an independently annotated sample quantify verbatim observations, uniquely reconstructable outputs, underdetermined but plausible reconstructions, and unsupported facts?"

1. Independent Source Fidelity Audit (Test 1)

To ensure that different source coverage intervals are thoroughly audited, we uniformly sampled 750 trajectories at random from each of the four coverage buckets above (3,000 trajectories and 9,970 action-observation steps in total), performed a 100.0% one-to-one exact pairing of each trajectory with its original web record, and weighted the bucket estimates by their exact proportions in the full pool to compute population-level estimates. An independent evaluator model from a different model family than the GLM-5.2 compiler (gpt-6.1-sol) compared each step against the original web record item by item under a blind setting with bucket information hidden.

To verify the error detection sensitivity and annotation standard reliability of the evaluator, we mixed 400 blind calibration sentinel samples into the review batches (including 300 trajectories with planted single-step factual contradictions: 100 with flipped success/failure polarity, 100 with injected unrelated infrastructure or error claims, and 100 with altered specific output numerical values, plus 100 unmodified clean controls), and simultaneously sampled 600 trajectories (2,000 observations in total, 150 trajectories per bucket) for item-by-item human verification against the original posts. On the blind sentinels, the evaluator accurately detected 96.7% of the planted factual errors on the modified steps (100.0% detection for polarity flips, 100.0% for injected error/infrastructure claims, and 90.0% for altered numerical outputs, with a 0.0% false positive rate on the target steps of unmodified clean controls). On the 2,000-observation stratified verification subset, the exact four-class agreement rate between the primary evaluator and item-by-item human verification reached 98.5%, and the binary agreement rate on whether an observation is source-grounded reached 99.0%.

Metric (SeedTraj pool-weighted; stratified sample N=3,000N = 3,000 trajectories, 9,9709,970 observations)[0.0, 0.1) (n=2,330n = 2,330 obs)[0.1, 0.4) (n=3,010n = 3,010 obs)[0.4, 0.7) (n=2,420n = 2,420 obs)[0.7, 1.0] (n=2,210n = 2,210 obs)Pool-Weighted Estimate
4-Way Observation Source Fidelity Breakdown
Verbatim copy of output shown in source (Verbatim)3.0%20.6%45.5%73.3%32.8%
Uniquely reconstructable from source record (Uniquely reconstructable)62.2%46.5%22.7%13.6%38.3%
Underdetermined but plausible reconstruction (Underdetermined plausible)27.9%24.9%24.8%9.0%22.3%
Contains unsupported or contradictory facts (Unsupported facts)6.9%8.0%7.0%4.1%6.6%
Total source-grounded or plausibly consistent93.1%92.0%93.0%95.9%93.4%

Across the full data pool, 93.4% of observations contain no unsupported facts: 32.8% are verbatim copies, 38.3% are uniquely reconstructable, and 22.3% reasonably fill in underdetermined runtime surface details while preserving the outcome established in the source post, leaving only 6.6% with unsupported or contradictory facts.

AuthorsOpenReview ↗

Response to Reviewer Dvvk (3)

Regarding "For any replayable subset, what fractions of commands, arguments, and outcomes agree with actual execution?"

2. Sandbox Execution Replay on a Self-Contained Subset (Test 2)

Most troubleshooting posts depend on external online services or specific operating system and hardware setups and therefore cannot be replayed, which is precisely the diversity that environment-first pipelines miss. We identified a replayable subset of 1,000 trajectories and replayed them sequentially in completely network-isolated Docker containers.

Execution Verification Metric (Replayable subset: N=1,000N = 1,000 trajectories, 3,0803,080 executed command steps)Agreement Rate with Actual Execution
Command agreement (matches record and executes validly)97.5%
Argument agreement (matches record and executes validly)96.8%
Outcome agreement (execution outcome polarity matches actual execution)92.4%

In this replayable subset, 97.5% of commands execute validly as recorded, 96.8% of arguments agree with actual execution requirements, and 92.4% of execution outcome polarities agree with actual execution.


For Weakness 1 (Part II) & Question 3

Weakness 1 (Part II): ArtifactStitch's semantic enrichment likewise does not establish compatible file contents, versions, paths, or preconditions. This limits the interpretation of the data as realistic executable experience, even if it is effective synthetic training material.

Question 3: What is the absolute correctness rate of ArtifactStitch edges and the rate of cross-task state contradictions? How does it compare with a strong semantic-neighbor-only composition baseline? The reported enrichment factor should not be interpreted as edge correctness.

Response:

  1. Design role and strict definition of the enrichment factor λ(a)\lambda(a): As defined in Section 2.3 of the main paper, ArtifactStitch preserves member trajectories that have passed Gates 1 to 3 copied byte for byte and places them under a unified merged tool declaration. Its supervision goal is to train agents to distinguish task boundaries, reuse relevant environmental context, and avoid erroneously carrying over outdated assumptions from completed tasks to the next request within a compatible shared working context (sharing files, tools, and domain vocabulary), rather than claiming that independent records from different web authors perform continuous in-place modifications on the exact same physical file version. We completely agree that λ(a)\lambda(a) serves solely as an artifact-level specificity filter (to remove generic placeholder filenames such as test.py and config.json) and should never be interpreted as an edge-level state correctness rate. Edge-level working-context correctness and state compatibility are directly provided by the independent blind audit below, and we have clarified this boundary in Appendix C of the revised paper.

  2. Independent blind audit comparing ArtifactStitch with a strong semantic-neighbor baseline (Semantic-Neighbor-Only): Based on the same Qwen3-Embedding-8B nearest-neighbor graph, we constructed a strong semantic-neighbor baseline (Semantic-Neighbor-Only) that strictly matches ArtifactStitch in scale, session length distribution, and monotonic workflow-stage constraints, without requiring shared filename artifacts. An independent blind audit by gpt-6.1-sol on 2,000 sampled sessions shows that ArtifactStitch achieves an edge-level shared working-context correctness rate (environment compatibility) of 89.8%, outperforming the strong semantic-neighbor baseline Semantic-Neighbor-Only (80.2%) by +9.6 percentage points, while reducing the cross-task hard state contradiction rate from 20.9% to 11.3%. Combining shared artifact anchoring, environment compatibility filtering, and enrichment testing allows ArtifactStitch to substantially improve context purity and reduce state conflicts compared to pure semantic neighbors. We have added this blind audit comparison to Appendix C of the revised paper.

  3. Matched-budget downstream training validation: The edge-level context purity gains measured in the blind audit are corroborated in controlled downstream training. Under the fixed 60B budget in Table 1 of the paper, directly comparing SeedTraj + ArtifactStitch against a session baseline that holds the exact same member trajectories fixed but concatenates them randomly (unrelated sessions):

    • Instruct SFT: The 8-benchmark average score improves from 32.7 to 34.2.
    • Think SFT: The 8-benchmark average score improves from 27.1 to 29.2.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.