acceptodds
Under review as a conference paper at ICLR 2027

Trained Agentic Context Management

Abstract

We study long context language models. Instead of training long context natively, or designing a long context harness, we train a model over the simplest possible harness: a tool to call itself with any specified prompt and a tool to read tokens in a range from the input context. We finetune Qwen3.6-35B-A3B on a diverse synthetic dataset using this harness. With only 8,000 tokens of context, our small model is as strong as GPT-5.4 with 1M tokens of context on the OOLONG-synth benchmark when document length exceeds 40K tokens.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLMs, long-context, multi-agent, SFT, synthetic data

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_saP2OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper fine-tunes Qwen3.6-35B-A3B to manage long contexts through a minimal harness with two tools, read_chunk and spawn_subagent. Each agent works under a small context budget, 8K tokens at evaluation. LoRA SFT on about 1.2K synthetic problems teaches three strategies: tree reduction, sequential folding and boundary handling. The model is trained only on 4–14K-token documents, yet it stays at or above 0.85 on RULER up to 320K tokens. On OOLONG-synth it is reported to match GPT-5.4 at long lengths. An RL stage was attempted without success.

Critical strengths

S1. The formulation is clean and minimal. Internalizing context decomposition through training, rather than hard-coding it into a harness, is a timely idea. The two-tool harness is simple and fully specified, which makes the setup easy to understand.

S2. The length-generalization result is notable and honestly reported. Although the model is trained on documents of at most 14K tokens, it operates on inputs of up to 320K tokens, and RULER NIAH stays at 1.00 throughout. Negative results are also reported candidly, including the CWE collapse and the unsuccessful RL stage.

Critical weaknesses

W1. The evaluation is too small, and the main claim is stated inconsistently. RULER uses 5 questions per task type and OOLONG-synth uses 30 problems per length. All results come from single runs without confidence intervals. At n = 30 the uncertainty is roughly ±15–18 pp, which is too large to support a claim of parity. The claim itself also varies: the abstract states parity with GPT-5.4 beyond 40K tokens, whereas Sec. 1.2 and Sec. 4 say beyond 80K. Reading approximately from Fig. 4, the model still trails GPT-5.4 at both 40K and 160K.

W2. Key baselines that would isolate the effect of training are missing. The paper reports neither base Qwen in the same harness nor a scripted harness that implements the same tree and fold strategies with base Qwen at the leaves. It also omits existing methods such as RLM and MemAgent, as well as base Qwen at its native context length beyond the 64K cap imposed by Tinker. Without the scripted-harness comparison in particular, it is unclear whether the gains come from training or from the hand-designed decomposition strategies encoded in the data.

W3. The training and evaluation distributions overlap substantially. The conclusion states that the authors "do not train on our evaluation distributions" (l. 354), yet several training tasks closely mirror the evaluation tasks: vt_novel resembles RULER VT, niah_* resembles NIAH, synth_topk resembles CWE/FWE, and labeled_records mirrors the construction of OOLONG-synth. Please confirm that the seven labelled datasets do not overlap with OOLONG-synth's sources, and evaluate on tasks with genuinely different structure.

W4. Data generation and inference cost are under-specified. It is unclear how the demonstrations were generated and verified, and how many training examples and tokens were used. Inference cost (number of subagents, total tokens, latency) and failure rates are not reported, which makes a fair comparison with a single GPT-5.4 call difficult.

W5. The study lacks depth, and the extent of automated research is unclear. The main text is about seven pages and contains relatively little experimental analysis, and the RL stage is described only anecdotally. According to the AI-use statement, generative AI tools were used at nearly every stage of the work. This includes writing all code, running experiments, creating figures, generating data, proposing hypotheses, designing the methodology and interpreting results. I appreciate this candid disclosure, and such use is permitted. Together with the limited depth of analysis, however, it gives the impression that the work may largely be the output of an automated research pipeline. A clearer account of how the authors verified the experiments, data and results, together with a deeper analysis of the learned behaviour, would increase confidence in the findings.

Questions for authors

Please see the Weaknesses section.

Reviewer_cY3eOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper fine-tunes Qwen3.6-35B-A3B to manage long documents using two tools: reading document chunks and spawning subagents. Synthetic demonstrations teach tree reduction, sequential folds, and boundary handling. Evaluation on RULER and OOLONG-synth shows that the model can process documents up to 320K tokens with an 8K context limit per agent.

Critical strengths
  1. The model demonstrates substantial length extrapolation: training documents are at most 14K tokens, while evaluation extends to 320K with an 8K per-agent context budget. This shows extrapolation beyond both the training document lengths and the context available to each agent.

  2. The method provides a simple interface and a concrete training recipe. Figures 1–3 clearly illustrate how the same two tools support aggregation, sequential processing, and chunk-boundary handling without fixing the decomposition workflow in the runtime harness.

Critical weaknesses
  1. The benefits of training and learned decomposition are not sufficiently isolated. Figure 4 compares the trained agentic model against models receiving the full document directly. Although the authors describe difficulties encountered by untrained models using the harness, they do not report quantitative results for this setting. A comparison with fixed tree-reduction/sequential-fold workflows is also missing. These controls would help quantify the benefit of SFT and establish whether learned decomposition improves on a prescribed workflow.

  2. The overall cost and performance-parity claim need stronger evidence. The 8K limit applies to each agent, but total token usage, agent calls, and latency are not reported, leaving the overall performance–cost tradeoff unclear. The GPT-5.4 parity claim also needs qualification: at 160K on OOLONG-synth, the proposed method scores 0.464, compared with approximately 0.56 for GPT-5.4 in Figure 4. With only 30 OOLONG-synth questions per length and no uncertainty estimates, the consistency of the claimed parity is difficult to assess.

Questions for authors
  1. How do base and fine-tuned Qwen compare using the same 8K agentic harness? For the fine-tuned checkpoint, how does learned decomposition compare with a fixed tree-reduction/sequential-fold workflow under comparable inference budgets?

  2. What are the total input/output tokens, subagent calls, and end-to-end latency at each document length? How do these compare with the full-document baselines?

  3. Can the authors provide exact scores and uncertainty estimates for the comparison with GPT-5.4 at 80K, 160K, and 320K, particularly given the gap at 160K in Figure 4? Please clarify where the parity claim holds and reconcile the different length thresholds stated in the abstract and main text.

Reviewer_McK5OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces a trained agentic context management framework that finetunes Qwen3.6-35B-A3B on a diverse synthetic dataset using a minimal two-tool harness (read_chunk and spawn_subagent). Operating within an 8,000-token per-agent context limit, the proposed small model achieves performance comparable to GPT-5.4 with a 1-million-token context window on the OOLONG-synth benchmark when document lengths exceed 40K tokens, while also scoring above 85% across document sizes from 10K to 320K tokens on the RULER benchmark.

Critical strengths
  1. High Efficiency under Resource Constraints: It enables a small model with an 8K token context budget to effectively match or approach the performance of massive frontier models (like GPT-5.4 with 1M tokens) on long-context tasks without requiring native long-context scaling or lossy token compaction.

  2. Scalable and Composable Decomposition Strategies: By avoiding hardcoded execution paths or static prompt structures, the model internalizes and dynamically composes flexible delegation strategies (such as tree reduce, sequential fold, and boundary handling) that scale gracefully to arbitrary document lengths.

Critical weaknesses
  1. Obviously, this is completly an AI-generated paper and the overall quality is low. Clear reject.
  2. Lack experiments on some widely used long context benchmarks such as LongBench v2, NIAH etc.
Questions for authors

Please write the paper to read like it was written by a human, and ensure the main text meets the page limit before submission.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.