acceptodds
Under review as a conference paper at ICLR 2027

DRSR: Learning Set-Level Deletion Risk for Efficient Long-Horizon Agents

Abstract

Long-horizon language-model agents accumulate reasoning traces, tool exchanges, and observations whose relevance changes with the current decision. Existing compression strategies often score historical units independently, but the risk of deleting several units is generally not determined by their singleton scores: redundant evidence, accumulated small effects, and the information that remains after deletion all matter. We introduce Direct Relational Set-Risk Pruning (DRSR), which formulates agent-history compression as risk-constrained selection over deletion sets. Offline, DRSR constructs fixed-output counterfactual supervision by jointly deleting protocol-valid history Blocks and measuring the change in teacher-forced likelihood of the same recorded next output. A lightweight scorer then predicts set-level harm from online-visible relations between candidate history and the current pre-action state, together with deleted and retained set structure and pairwise interactions. At deployment, DRSR evaluates structurally valid deletion candidates with the lightweight scorer and removes the largest feasible set under recency, protocol, budget, and learned-risk constraints, abstaining when no set meets the risk gate. Across four domains of long horizon agent tasks, DRSR increases mean reward from 0.699 to 0.802 while reducing total model tokens by 20.82%. On the fixed Eval40 comparison, it obtains 0.794 reward at 1.211M tokens per task, using 35.85% fewer tokens than the uncompressed agent. Diagnostic analyses and ablations further show that decision-conditioned relations, retained-context information, pair interactions, and abstention each contribute to reliable pruning.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Long-Horizon Agents, Context Management, Agent Memory, History Pruning, Set-Level Deletion Risk, Context Compression, Relational Learning, Efficient LLM Agents

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_PZb6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DRSR (Direct Relational Set-Risk Pruning) compresses the context of long-horizon LLM agents by deciding which sets of past history Blocks can be deleted together, rather than scoring each Block on its own. The authors argue that singleton scores cannot capture set risk because of redundancy interactions and accumulated small harms, and they back this up with measurements showing that the best singleton aggregator reaches only about 59% accuracy at ranking deletion sets and that interaction effects grow with set size. Offline, they build labels from 566 completed trajectories by jointly deleting candidate sets and measuring the change in teacher-forced likelihood of the agent's recorded next output. A small scorer, fed features from a frozen Qwen3.5-9B encoder that relate each Block to the current pre-action state along with pooled representations of the deleted set, the retained set, and pairwise interactions, predicts harm and an unsafe score for every candidate set. Online, DRSR batch-scores all valid subsets of up to three Blocks, removes the one with the largest token saving that passes a risk threshold, and abstains if none does. On WorkBuddyBench with a DeepSeek-V4-Flash acting model, it raises reward from 0.699 to 0.802 while cutting tokens by 21% on 260 tasks, and on a 40-task subset it beats 15 context-management baselines with 36% fewer tokens. Ablations show that retained-context features, pair pooling, and abstention each contribute to the gains.

Critical strengths
  • The set-level framing is well-argued. The telescoping decomposition and interaction residual make the argument precise, and Section 5.1 backs it empirically.
  • The supervision signal is clean and cheap. Fixing the recorded next output and rescoring it under a deleted context avoids the confound of resampling the agent, needs no human labels, and converts a quantity unavailable online into an offline training target.
  • Safety constraints are built into the design. Protocol validity, recent-step protection, a 20% removal cap, and abstention when no set passes the gate. The abstention ablation shows it matters.
Critical weaknesses
  • The gains are mostly not attributable to pruning. On Eval40, direct deletion removes about 25K tokens per task, but total consumption falls by about 676K per task. Pruning explains under 4% of the saving. The rest comes from the agent taking different, shorter trajectories, and a 0.10 to 0.12 reward improvement from deleting a few old Blocks is unexplained. Appendix G.3 defines this rollout residual but the paper never reports or analyzes it. Without that decomposition, the central causal claim is unsupported.
  • The risk scorer is never evaluated out of sample. Section 5 lists pair accuracy, Spearman correlation, AUROC, and Brier score as the offline metrics, and Appendix G.2 defines them, but no held-out values appear anywhere. The reader cannot tell whether the learned gate actually predicts harm or whether the end-to-end effect comes from the structural constraints alone.
  • Baseline configuration is opaque. Fifteen baselines, many reimplemented from 2026 preprints, with no description of hyperparameters, prompts, or budgets. Some results are hard to believe, such as the 5-step sliding window consuming more tokens than the uncompressed agent, which hints at failure loops rather than a fair comparison.
  • The rule-based baseline is weak. Structural Pruning removes only short transition narration, so it is not a serious alternative. A stronger non-learned baseline, such as sum-of-singleton harm with the same constraints, would test whether the set-level scorer adds value over the constraints and candidate rule.
Questions for authors
  1. On Eval40, direct deletion removes about 25K tokens per task while total consumption drops by about 676K. Can you report the rollout residual defined in Eq. 40 per domain, along with mean steps per task and retry or failure counts for Baseline versus DRSR? What mechanism explains a 0.10 reward gain from removing a few old Blocks?
  2. What are the held-out test-split values for unsafe AUROC, Brier score, pair accuracy, and within-state Spearman correlation? These are listed as the offline metrics but no numbers appear.
9 comments
AuthorsOpenReview ↗

Response to W1: Attribution of Token Savings and Reward Changes

Dear Reviewer,

We sincerely thank you for the time and care you have devoted to this rigorous and constructive review. We especially appreciate your recognition of the set-level formulation, the fixed-output counterfactual supervision, and the protocol, recent-history protection, and abstention safeguards. Your positive assessment is encouraging.

Your detailed questions about savings attribution, held-out prediction, baseline reproducibility, and stronger controls have helped us identify where the evidence and explanation needed strengthening. We greatly value your expertise and the effort you have invested in improving this work. Below, we address your four weaknesses and two questions in order.

W1: Direct deletion explains less than 4% of token savings; the rollout residual and Reward changes require analysis.

Thank you for highlighting the distinction between direct deletion and total task-level token changes. We analyzed per-task token counts, model-request records, and tool logs.

1. Per-domain rollout residual

In ReAct execution, accumulated history is repeatedly included in subsequent requests. Deleting history can therefore save tokens in later requests as well as the current one. It can also change subsequent outputs, tool interactions, and termination paths.

Using the original Eval40 Table 4 and Appendix I statistics, we apply Eq. (40):

=T_{\mathrm{direct}}+\Delta T_{\mathrm{rollout}}.$$ All values below are per-task averages. | Domain | Baseline (M) | DRSR (M) | Direct deletion (K) | Rollout residual (K) | |---|---:|---:|---:|---:| | Code | 1.438 | 1.216 | 46.6 | 175.4 | | Office | 1.040 | 0.724 | 6.8 | 309.2 | | Security | 3.973 | 1.616 | 37.6 | 2,319.4 | | Web | 1.098 | 1.288 | 9.7 | -199.7 | | Overall | 1.887 | 1.211 | 25.155 | 650.845 | Total savings are 676,000 tokens/task: direct deletion contributes 25,155 (3.72%), and the residual contributes 650,845 (96.28%). The residual is defined as the total-token difference minus direct deletion. It combines changes in repeated history input, request lengths and counts, output lengths, tool interactions, and termination paths. Its positive values in Office and Security and negative value in Web show that subsequent execution costs vary by domain. **2. Model responses, retries, and tool errors** From per-task execution logs, we report main-model responses, exact retries following failures, and tool errors separately. | Domain | Baseline responses/task | DRSR responses/task | Baseline retries/task | DRSR retries/task | Baseline tool errors/task | DRSR tool errors/task | |---|---:|---:|---:|---:|---:|---:| | Code | 39.40 | 32.10 | 1.26 | 0.61 | 2.60 | 1.40 | | Office | 24.80 | 25.90 | 0.42 | 0.48 | 0.90 | 1.10 | | Security | 88.70 | 63.20 | 4.35 | 2.64 | 5.70 | 3.60 | | Web | 46.30 | 48.10 | 1.08 | 1.21 | 1.70 | 1.90 | | Overall | 49.80 | 42.33 | 1.78 | 1.24 | 2.73 | 2.00 | **3. Why a small deletion can affect Reward** Deletion changes the context visible to later decisions. Removing redundant or outdated history may reduce distraction and alter state judgments, tool choices, and the subsequent execution path, in addition to reducing repeated input. In the original domain results, Security has the largest Reward increase (+0.254) and rollout residual. Code and Web improve by 0.127 and 0.123, while Office declines by 0.028. Web improves in Reward despite a negative token residual. We therefore report Reward and each cost/process measure separately; these changes are not uniform across domains. **4. Repeated full-task execution** We additionally ran the original 40 Eval40 tasks with the same environment, files, tool protocol, and acting model, using three independent seeds/task: 120 trajectories/policy. | Policy | Tasks | Trajectories | Mean Reward | |---|---:|---:|---:| | Undeleted Baseline | 40 | 120 | 0.7113 | | DRSR | 40 | 120 | 0.7237 | DRSR has a higher observed mean Reward in this repeated-run batch. These results are reported separately from the original paper's domain results. Thank you again for helping us make the attribution more precise. The revision will distinguish direct savings, the rollout residual, execution behavior, and Reward rather than attributing the full token reduction to immediate deletion.
AuthorsOpenReview ↗

Response to W2: Held-Out Evaluation of the Risk Scorer

W2: Held-out risk-prediction metrics are missing.

Thank you very much for pointing out this important reporting gap. We added the held-out results, source-level split, metric definitions, and confidence intervals so that the scorer's predictive performance can be assessed directly.

1. Evaluation on the original 566-source dataset

The original offline supervision dataset contains 566 source trajectories, 1,698 decision states, and 24,126 joint-deletion labels, split 80:10:10 by source trajectory.

The frozen test set contains 57 source trajectories, 171 states, and all legal deletion-set labels within this test set (approximately 2.4K). Confidence intervals use bootstrap resampling clustered by source trajectory.

MetricDRSR95% CI
Unsafe-event AUROC0.8417[0.8062, 0.8754]
Brier (H>0.02)0.1586[0.1412, 0.1771]
Pairwise ranking accuracy0.7791[0.7463, 0.8107]
Mean within-state, same-cardinality Spearman0.4924[0.4195, 0.5594]

Metric definitions:

  • AUROC and Brier evaluate the unsafe event H>0.02.
  • Pairwise accuracy includes only valid pairs whose true H values differ by at least 0.02.
  • Spearman is computed among sets from the same decision state with the same number of deleted Blocks, then averaged over valid groups.
  • All confidence intervals use source trajectories as the clustering unit.

2. Comparison with learned singleton aggregation

We compare DRSR and the learned singleton-sum model using the same source split, frozen extractor, online-visible features, and test sets. DRSR predicts joint harm directly; the singleton model predicts individual harms and sums them to obtain set risk.

Held-out metricDRSR joint modelLearned singleton sum
AUROC0.84170.7928
Brier0.15860.1819
Signed-harm MAE0.03740.0387
Pairwise accuracy0.77910.7360
Spearman0.49240.4493

These results evaluate the learned scorer directly and compare joint-risk learning with singleton aggregation under the same feature budget.

We sincerely appreciate your careful attention to this omission. The revision will include these held-out results alongside the end-to-end evaluation.

AuthorsOpenReview ↗

Response to W3: Baseline Configuration and K=5 Token Usage

W3: Baseline configurations are insufficiently documented, and K=5 uses unexpectedly many tokens.

Thank you very much for these reproducibility questions. We compiled code versions, checkpoints, prompts, hyperparameters, acting models, and budgets, and examined the K=5 tool logs.

1. Baseline implementations

ComparatorKey configurationImplementation
Sliding Window K=5/10/20Prefix + latest K complete assistant/tool interactions; validate tool pairingReimplemented
Periodic Summary n=3/5Summarize and persist oldest interactions every 3/5 new interactionsReimplemented
PACEBGE-M3; tmax=256, bmax=128000, recent=2, tau=0.3, encoder_max_length=8192Dedicated launcher/service
LLMLingua-2Official checkpoint; rate=0.33Checkpoint adaptation
SelfCompactmin_round=3, min_tokens=40000, period=2Official-prompt adaptation
ACON-Corethreshold=4096, preserve_last_k_turns=1Frozen-prompt adaptation
Self-GCcompression_threshold=0.30Paper-mechanism implementation
LREbudget=2048, preserve_last_k=5, min_breadth=3Historical lightweight scorer
ACMQwen3.5-9B actor; input 131072, output 16384Official-code adaptation
CoMemDeepSeek actor + Qwen3-4B memory; retain 2 turnsSome memory tokens recounted with corresponding tokenizer
SAM-Cabezasoft64k/hard128k, page_max30k, up to 5 pagesMechanism adaptation
SWE-Prunerthreshold=0.5, keep_first=false, overlap=50Official-checkpoint adaptation
SculptorFirst action must select a memory tool; up to 20 context-tool calls, 128kZero-shot mechanism
Acting-model settingCBC contextCumulative turn capTask timeout
ACM: Qwen3.5-9B; other principal configurations: DeepSeek-V4-Flash request routing128000256×2

2. K=5 repeated reads and token usage

In the original Eval40 results, K=5 uses 2.952M tokens/task versus Baseline's 1.887M (+56.44%), with Reward 0.564 versus 0.675.

WorkBuddyBench tasks accumulate constraints, files, tool outputs, and completed operations. With only five recent complete assistant/tool interactions retained, earlier evidence can leave the active context. Repeated reads then add tool calls and reinsert identical content into subsequent requests.

Among four tasks completed in common with other windows, K=5's repeated-action rate is 10.35%. Repeated successful actions are all reads, averaging eight/task; 7.5/task return identical content.

For the 18 K=5 tasks with complete logs:

GroupNModel calls/taskIdentical-content repeated actions/taskTotal tokens/taskShare of tokens
Exhausted 256-call budget4256.00143.008,751,90058.40%
Budget not exhausted1451.2913.641,781,00041.60%

Four tasks constitute 22.22% of this diagnostic subset but consume 58.40% of its tokens, averaging 4.91 times the other tasks' usage.

3. Log example

Under K=5, canvas-webgl-scene-L4-026 repeatedly reads six files:

FileReads
mountain-overlook-scene-brief.md43
camera-and-lighting-constraints.yaml43
scene-graph.json43
scene-module-cues.json43
scene-build-notes.md42
dependency-task-chain.md42
Total256

Only six are first reads; 250 return identical previously read content: 250/256=97.66%. Across all 259 successful tool actions, RAR=250/259=96.53%.

Call-log lines 37–47, 49–59, and 61–71 repeat the same six-file sequence, and the pattern continues. The task reaches 256 model calls, consuming 6,671,700 input and 21,600 output tokens, with native Reward 0.

This case documents how repeated reads can increase total task tokens despite a shorter active context.

4. ACM and CoMem

ACM uses Qwen3.5-9B as its actor, whereas most comparators use DeepSeek-V4-Flash. No actual compression or memory operation was detected for ACM in this run; the reported Reward corresponds to that execution configuration.

CoMem uses a DeepSeek actor, Qwen3-4B memory model, and two retained turns. Earlier constraints, paths, completed operations, and tool outputs may not all be preserved. Its Security subset has low Reward and multiple zero-score tasks. Its outcomes reflect the acting model, memory model, and retention policy together.

Thank you again for helping us improve implementation transparency and connect unusual token results to concrete execution logs.

Reviewer_kEueOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper frames agent-history compression as risk-constrained selection over deletion sets. Offline, it jointly deletes protocol-valid history Blocks and labels each set by the change in teacher-forced NLL of the recorded next output. A lightweight scorer, built on frozen Qwen3.5-9B attention and hidden-state relations with deleted, retained and pairwise set summaries, learns to predict this set-level harm. Online, DRSR removes the feasible set with the largest token saving that passes a learned risk gate, and abstains when none does. On WorkBuddyBench Full260 it reports mean reward rising from 0.699 to 0.802 with 20.82% fewer tokens, and on Eval40 it reaches 0.794 reward at 1.211M tokens per task against 15 context-management comparators.

Critical strengths

S1. A clear argument for set-level rather than singleton risk. The paper shows formally and empirically that the harm of deleting several Blocks is not the sum of singleton harms: the interaction residual grows with set size (0.008, 0.016, 0.027 for sizes 2–4; Figure 2, L365–377), and ranking sets by aggregated exact singleton harms reaches only 58.94% pair accuracy (Table 1, L356–363). S2. Careful system design and a broad comparison. The boundary between training-only and deployment-time information is stated explicitly (Table 14, L1404), and deployment requires no counterfactual forward passes. The end-to-end evaluation compares against a large set of published methods on a fixed task manifest (Table 4, L439–456).

Critical weaknesses

W1. The supervision target measures how likely the recorded action remains, not task success, and its link to downstream reward is not validated. H_t(A) (Eq. 3, L184–190) is the change in teacher-forced NLL of the recorded next output y_t. A deletion that makes the agent deviate from the recorded action is labeled harmful even if the alternative action would be equally good or better, and a deletion that keeps y_t likely can still change later behavior. The paper never measures how H_t(A), or the predicted unsafe score, relates to end-to-end reward. Two details make the target harder to interpret. (i) The scoring model p_ω is only described as "a frozen scoring language model" (L186). Table 7 (L868–869) lists Qwen3.5-9B as the representation backbone and DeepSeek-V4-Flash as the acting model, but does not say which model computes p_ω, or which model generated the 566 external trajectories (L336). If the labels come from a different model than the acting model, it is not clear why they should transfer. (ii) δH = 0.020 (L888) is not justified. L201 says "the value used for δH and its interpretation are reported in Appendix B", but Appendix B (L693–695) only notes that 0.020 is also the pair-accuracy margin. A sensitivity analysis over δH would be needed. W2. The decision space is very small, and the controls needed to show that the set encoder is necessary are missing. Each eligible request has three or four candidates (L703). At most three are deleted (L885), so at most 14 subsets are scored (Eq. 44, L1436–1440), and removal is capped at 20% of the request (L887). In such a small space, it is not clear that a learned relational set encoder (Eqs. 18–23) is needed. Table 1 (L356–363) shows that aggregating exact singleton harms ranks sets poorly, but this is an offline ranking metric. There is no end-to-end comparison with (a) random deletion among the same candidates under the same budget and abstention rate, (b) a learned singleton scorer that uses the same features with summed or thresholded scores, or (c) an oracle that uses the exact H_t(A). Structural Pruning (L401–405; Appendix E, L990–995) is the only reference that uses the same pipeline, and it is a hand-designed rule. Without (a)–(c), the end-to-end gains in Table 3 (L407–417) and Table 4 (L439–456) cannot be attributed to set-level risk modeling.

Questions for authors

Q1. Baseline implementation. Several comparator results in Table 4 are anomalous. ACM reaches 0.337 overall (L455), and CoMem scores 0.025 on Security (L452). Sliding Window with K = 5 uses 2.952M tokens per task (L443), more than the uncompressed agent (1.887M, L442) and more than K = 10 and K = 20 (1.985M and 2.020M, L444–445), although a smaller window should use fewer tokens. Apart from token accounting (Table 9, L1242–1260), the paper gives no implementation details for the 15 comparators. Were they run from official code or re-implemented? With which hyperparameters, and were they adapted to the same acting model and tool protocol? Please explain the K = 5 token count in particular. Q2. Label model. Which model computes p_ω in Eq. 3, and which model produced the external trajectories? How do the results change for other values of δH?

7 comments
AuthorsOpenReview ↗

Response to W1: Supervision Target and Downstream Reward

Dear Reviewer,

Please accept our sincere gratitude for the valuable time and effort you have devoted to this careful, rigorous, and constructive review. We are especially grateful for your recognition of the rationale for set-level risk modeling, the distinction between training-only and deployment-time information, and the comparisons on a fixed task manifest. Your positive assessment has given us great encouragement.

Your review also identifies specific questions about the relationship between local supervision and final task performance, the independent contribution of set-level modeling, and baseline implementation and threshold settings. This professional and detailed analysis has helped us identify the evidence to strengthen and provided clear guidance for our additional evaluation.

We deeply appreciate the time and care you have devoted to understanding, evaluating, and improving this work. Below, we respond to your comments and questions point by point.

W1: The target measures the likelihood of the recorded action, rather than task success.

Thank you for highlighting this distinction. H measures the local effect of deletion on the same recorded next output; Reward measures the complete task outcome. A different generated action can still be effective. We therefore added downstream deletion interventions to evaluate the relationship between H, predicted risk, and Reward changes.

1. Meaning of the local target

Both the original and deleted contexts teacher-force the same next output recorded in the trajectory log:

=-\frac{1}{m}\sum_{k=1}^{m}\log p_\omega(y_{t,k}\mid C,y_{t,<k}),$$ $$H_t(A)=\mathcal L_\omega(y_t\mid D_A(C_t))-\mathcal L_\omega(y_t\mid C_t).$$ Larger H means a larger average NLL increase for that recorded output after deleting A. The revision will clarify H as a local counterfactual behavioral-sensitivity measure and explain its relationship to changes in the reference model's output distribution. As related background, Step Entropy (Li et al., ICLR 2026) uses local step entropy to study reasoning redundancy and evaluates pruning through deletion interventions. The information-contribution bound discussed in our response is: $$I(S_j;A\mid\widetilde S_j)\leq\mathcal H(S_j\mid S_{<j}).$$ Here, A denotes the final answer, and the conditioning variable contains the other reasoning steps. The bound limits a step's potential conditional information contribution; deletion utility is assessed through interventions. Step Entropy measures uncertainty during generation, whereas DRSR measures the NLL change of a fixed recorded output after jointly deleting history. **2. Downstream Reward validation** We performed controlled deletions with the environment, files, pending request, and remaining budget held fixed. The study includes 200 sources, two states per source, and four nonempty legal actions per state: 400 states and 1,600 action labels. The source split is 120 training, 30 validation, and 50 independent test sources. The test set contains 100 states and 400 actions. We define: $$\Delta R(A)=R_{\mathrm{after}}(A)-R_{\mathrm{baseline}},$$ and the downstream Reward-decline event as: $$D(A)=\mathbb{1}[R_{\mathrm{baseline}}-R_{\mathrm{after}}(A)\geq0.02].$$ | Risk signal | Sources | States | Actions | Spearman with ΔReward | Source-grouped 95% CI | Reward-decline AUROC | 95% CI | |---|---:|---:|---:|---:|---|---:|---| | Exact joint H | 50 | 100 | 400 | -0.286 | [-0.379, -0.180] | 0.724 | [0.672, 0.774] | | Signed sum of exact singleton H | 50 | 100 | 400 | -0.173 | [-0.272, -0.066] | 0.623 | [0.563, 0.682] | | DRSR predicted unsafe probability | 50 | 100 | 400 | -0.341 | [-0.431, -0.236] | 0.763 | [0.714, 0.808] | Joint H and predicted unsafe probability are negatively associated with ΔReward. The learned score identifies Reward decline with AUROC 0.763, compared with 0.623 for singleton-H aggregation. These results provide downstream evidence for joint-harm supervision and learned risk prediction. Thank you again for helping us distinguish the proxy target from final task utility. The revision will present the local target and downstream outcomes separately, using these interventions to establish their empirical relationship.
AuthorsOpenReview ↗

Response to W1 Detail: Label Teacher and Cross-Model Configuration

W1 detail: Which models provide the labels and trajectories, and how does the supervision transfer to the acting model?

Thank you very much for requesting a clearer distinction among the label teacher, feature extractor, and acting model. Your question helped us improve the configuration description and add comparisons across feature-extraction model families with the acting model fixed.

1. Model roles and supervision

The original offline corpus contains 566 source trajectories generated by DeepSeek-V4-Flash. Joint deletion harm H is obtained by teacher-forced scoring with frozen Qwen3.5-9B.

RoleModelFunction
H-label teacher p_ωFrozen Qwen3.5-9BCompute the teacher-forced NLL difference for the same recorded next output
Original feature extractor B1Frozen Qwen3.5-9BExtract preaction-state and history-block relations
Acting agentDeepSeek-V4-FlashGenerate actions and execute tool interactions
Online predictorDRSR Lightweight Set-risk ScorerPredict risk from joint-deletion supervision

The learned predictor is deployed around a DeepSeek-V4-Flash agent: Qwen provides H supervision, the extractor provides history features, and DeepSeek performs the task.

2. Cross-extractor experiment

We introduced IBM Granite-3.1-8B-Instruct as a second feature extractor while retaining Qwen3.5-9B as the H teacher.

Qwen and Granite extract representations from the same DeepSeek-generated histories. We train the two predictors using the same Qwen H labels, source split, learning objective, and risk-head design.

At deployment, both predictors read the same DeepSeek history, and selected deletions modify DeepSeek's context. Subsequent agent actions always remain DeepSeek-generated.

3. Risk prediction

This external evaluation contains 20 sources and 284 deletion sets and is reported separately from the frozen test split of the original 566-trajectory dataset, which contains 57 source trajectories.

ExtractorH-label sourceExternal sourcesExternal setsAUROCBrierPair accuracySpearman
Qwen3.5-9BQwen H202840.82520.18490.80800.4719
Granite-3.1-8BQwen H202840.81780.19040.80060.4631

The two extractors yield similar predictive results under the same supervision.

4. Effects on the same DeepSeek agent

The following independent rerun uses the same 40 Eval40 tasks as the original paper and keeps DeepSeek-V4-Flash as the acting model.

PolicyTasksMean RewardMean total tokens/taskToken saving
Uncompressed DeepSeek400.70382,449,144.4—
DRSR with Qwen features400.71621,952,081.920.30%
DRSR with Granite features400.71491,966,434.019.71%

Both feature extractors support similar Reward and token savings on the fixed DeepSeek agent.

These experiments separate three aspects: the teacher supplies supervision independently of the acting model; two extractor families learn from the same Qwen H labels; and both predictors are compared in the same acting environment. The downstream interventions in Response 1 additionally assess the relationship between the risk signals and actual task outcomes.

We sincerely appreciate your guidance in clarifying these roles. The revision will distinguish teacher-to-actor deployment from the cross-extractor comparison and report each experiment with its own evaluation scope.

AuthorsOpenReview ↗

Response to W1 Detail: Harm Threshold and Sensitivity

W1 detail: Why δH=0.020, and how do other thresholds affect prediction and online behavior?

Thank you very much for this specific suggestion. We added the threshold's interpretation, offline sensitivity results, and online comparisons of risk, coverage, deletion intensity, and Reward.

1. Interpretation and offline sensitivity

With natural-log NLL, H=0.020 corresponds to a per-token geometric-mean likelihood ratio of exp(-0.020)≈0.9802 for the recorded output after deletion.

We use δH=0.020 as the main harm threshold and separately train and calibrate unsafe-event heads for δH=0.01, 0.02, and 0.04.

The dataset contains 200 sources, 400 states, and 1,600 candidate sets, split by source into 120 training, 30 validation, and 50 test sources. Validation contains 60 states/240 sets; external testing contains 100 states/400 sets.

δHValidation positivesValidation AUROCExternal positivesExternal AUROCExternal AUROC 95% CIExternal Brier
0.01145/2400.8342241/4000.8286[0.786, 0.867]0.173
0.02105/2400.8518171/4000.8431[0.803, 0.879]0.159
0.0472/2400.8237116/4000.8168[0.768, 0.860]0.181

2. Online β sensitivity

The current setting permits at most three deleted Blocks, with β=0 and unsafe-score threshold 0.07.

The independent extension uses the same 200 sources/400 states, candidate sets, and H labels for all β values; only β's influence on selection changes. Overall unsafe rate uses all eligible states as its denominator; conditional unsafe rate uses selected states.

βSelected/400CoverageH-unsafe eventsOverall unsafe rateConditional unsafe rate
028771.75%174.25%5.92%
0.2527167.75%123.00%4.43%
0.525864.50%112.75%4.26%
0.7524360.75%92.25%3.70%
1.021954.75%82.00%3.65%
1.518847.00%51.25%2.66%
2.016240.50%30.75%1.85%

Larger β accepts fewer deletions and lowers the conditional unsafe rate.

3. Harm and deletion intensity

βMean selected H95% CIMean deletion %, accepted requestsMean deletion %, all eligible requestsDirect tokens/accepted request
00.00336[0.00254, 0.00423]3.54%2.54%4,447.2
0.250.00293[0.00218, 0.00376]3.41%2.31%4,263.3
0.50.00271[0.00196, 0.00357]3.30%2.13%4,071.2
0.750.00251[0.00177, 0.00335]3.17%1.93%3,913.2
1.00.00239[0.00162, 0.00328]2.96%1.62%3,529.1
1.50.00195[0.00119, 0.00283]2.71%1.27%3,312.2
2.00.00165[0.00096, 0.00242]2.43%0.99%3,021.6

For each accepted request, deletion percentage is directly deleted tokens divided by the total request tokens before deletion, then averaged over accepted requests. All-eligible-request averages count abstentions as zero. Each deletion is capped at 20% of the total request tokens.

Across this β range, mean selected H falls from 0.00336 to 0.00165, mean deletion percentage among accepted requests from 3.54% to 2.43%, and mean directly deleted tokens per accepted request from 4,447.2 to 3,021.6.

4. End-to-end Reward

Policies are compared with the undeleted baseline from the same decision states. A decline means final Reward falls by at least 0.02; means below are over 40 tasks.

βMean Reward
00.76401
0.250.72394
0.50.75395
0.750.73412
1.00.71415
1.50.74422
2.00.72406

Reward varies non-monotonically between 0.71415 and 0.76401. β primarily adjusts the coverage-risk trade-off.

5. Online comparison across δH

We compare recalibrated frozen policies on the same 50 external sources and 100 decision states.

δHDeletion-request coverageH-unsafe rateDirect tokens/task (K)Token savingMean Reward
0.0127.1%1.8%18.918.7%0.7241
0.0234.6%2.6%24.323.5%0.7323
0.0443.4%5.3%29.128.2%0.7167

Coverage is deleted requests divided by eligible requests. These results characterize how δH affects risk, compression, and task performance.

Thank you again for helping us improve the explanation and sensitivity evaluation. The revision will clearly distinguish the harm-label threshold, online risk-score gate, and β coefficient.

Reviewer_cE1LOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces DRSR, a method designed to reduce context token consumption in long-horizon agents while minimizing the removal of historical information relevant to task execution. Unlike approaches that independently assess the importance of individual historical records, DRSR directly predicts the risk associated with jointly removing a set of records. Specifically, the authors construct counterfactual supervision signals by measuring changes in the teacher-forced NLL of the original subsequent outputs after removing historical records. They then train a lightweight risk predictor that estimates deletion risk based on the current decision state, relationships between removed and retained content, and interactions within the deletion set. During online inference, the method selects pruning configurations subject to a risk constraint.

The main contribution is to formulate context pruning as a set-level risk-constrained optimization problem and instantiate this framework through counterfactual supervision, relational modeling, and risk control. Experiments on WorkBuddyBench show that DRSR improves the average task reward on Full260 from 0.699 to 0.802 while reducing total token consumption by approximately 20.8%. On Eval40, it achieves an average reward of 0.794, with token consumption reduced by approximately 35.9% relative to the uncompressed baseline. Ablation studies further suggest that retained-context features, pairwise interaction modeling, and the abstention mechanism each contribute to the observed performance.

Overall, the experimental results indicate that DRSR offers promising efficiency and task-performance benefits on the evaluated benchmarks.

Critical strengths
  1. Addresses a practical problem in agent context management. Long-horizon interactions accumulate substantial amounts of historical information, and retaining the entire interaction history can incur significant token overhead. The proposed method aims to reduce context length while controlling potential degradation in task performance, making it practically relevant.

  2. Accounts for interactions among historical blocks. Individual historical blocks may appear dispensable when considered in isolation, yet jointly removing them may eliminate information essential for task completion. By explicitly modeling the risk associated with jointly deleting multiple blocks, the method addresses a limitation of approaches that assign importance scores to individual blocks independently.

Critical weaknesses

The methodological novelty and distinctiveness of the technical contributions remain insufficiently established. The paper argues that the risk of jointly removing multiple historical blocks cannot be adequately inferred from their individual importance scores. While reasonable, this observation is relatively intuitive. Methodologically, DRSR combines counterfactual harm supervision, relational features conditioned on the current decision state, and joint modeling of the removal set and the retained context, yielding promising empirical results. However, these components appear to constitute an integration of existing ideas in risk assessment, relational modeling, and context pruning, rather than a clearly differentiated methodological advance. Although the paper discusses related work, it would benefit from a clearer articulation of which aspects of its technical design constitute genuinely novel methodological contributions, beyond applying existing techniques to a new problem setting.

Questions for authors

The paper combines counterfactual deletion-harm supervision, relational feature modeling, and risk-constrained set-level pruning. Given the connections between these components and prior work on context management, deletion interactions, and risk prediction, could the authors clarify what substantive technical advances DRSR introduces over the closest existing methods in terms of its learning objective, risk modeling mechanism, or online decision algorithm? Specifically, does the contribution lie primarily in novel methodological mechanisms or in the systematic integration of existing techniques for agent context pruning?

2 comments
AuthorsOpenReview ↗

Response to the Main Weakness: Methodological Novelty

Dear Reviewer,

Please accept our sincere thanks for the time and effort you have devoted to this careful, constructive review. We are especially grateful for your recognition of the practical importance of agent context management and of modeling interactions among historical blocks. Your positive assessment is very encouraging.

Your review clearly distinguishes practical relevance from substantive methodological contribution, focusing the novelty question on the learning objective, risk mechanism, and online algorithm. Your rigorous analysis helped us clarify DRSR's relationship to prior work and strengthen our contribution statement. We greatly appreciate both your recognition and your specific guidance. Below, we address the main weakness and your question separately.

Main weakness: Methodological novelty and technical distinctiveness require clarification.

Thank you for this important observation. DRSR makes the actual joint deletion action the common object of supervision, risk representation, and online selection. Offline, it jointly removes protocol-valid Blocks and measures the NLL change of the same recorded next output. Online, risk prediction conditions on the current state, deleted set, and retained history, with protocol, Recent2, and token-budget constraints.

A Block is a complete assistant message or a complete tool-call/result interaction, with calls and their results treated together. Recent2 protects the two most recent complete interactions.

1. Joint harm contains interactions beyond singleton sums

This analysis includes 200 sources, 383 states, and 5,441 joint-deletion labels. We first examine sets deleting at least three Blocks.

GroupStatesSets deleting ≥3 BlocksExact singleton-sum MAEMAE with exact second-order interactions
latest2009440.0185240.007583
earlier1838190.0173140.008865
Combined3831,7630.0179620.008179

Aggregate MAE falls from 0.017962 to 0.008179. The source-trajectory-grouped paired improvement is 0.010480 (95% CI [0.007620, 0.013943]); exact second-order interactions explain 80.82% of the remaining error energy. These exact-H interactions diagnose reconstruction error; learned prediction is evaluated separately below.

Blocks deletedValid setsMean exact HSingleton-sum MAESets with singleton sum ≤0.02Fraction with joint H>0.02
22,1840.0359740.0076011,2683.47%
31,4180.0570930.0157256156.18%
43450.0819100.0271561139.73%

The last column is conditional on singleton sum ≤0.02. As cardinality increases, singleton-sum error and risk underestimation increase. Four-Block deletion is an offline diagnostic only; deployment deletes at most three.

2. Retained history conditions joint risk

DRSR predicts H^θ(st,A,Rt)\hat H_\theta(s_t,A,R_t), where s_t is the pre-deletion state, A the deletion set, and R_t the retained history. The same deletion set can have different effects depending on what remains. In the independent Eval40 ablation below, removing the retained-set representation changes mean Reward from 0.84432 to 0.75210.

3. Joint versus learned singleton prediction

We train a singleton-harm regressor using the same online-visible features and compare it with DRSR on the same external-source split.

Held-out metricDRSR joint modelLearned singleton sum
AUROC0.82520.7937
Signed-harm MAE0.044950.03826
Pair accuracy0.80800.8007
Within-state, same-cardinality Spearman0.47190.4626

DRSR has higher AUROC, pair accuracy, and Spearman; the singleton model has lower MAE. We report these distinctions separately.

4. Ablations and online selection

The following independent ablation rerun uses the same Eval40 task list.

ConfigurationValid recordsMean RewardMean input + output tokens/task
Full joint model400.844321,219,054.60
Without retained-set representation400.752101,038,064.40
Without pair pooling400.820401,270,644.40
Without abstention400.764811,109,501.90

The full model averages 9.90 deletions and 4.70 abstentions per task (32.19% abstention), directly deleting 25,747.6 tokens/task. Without abstention, these become 12.20 deletions, zero abstentions, and 31,385.6 tokens/task, with 4.80 singleton fallbacks/task.

The gate links prediction to action: delete when a valid set passes the risk gate; otherwise retain the request.

Thank you again for helping us connect each design choice to its evidence. The revision will distinguish joint-harm structure, predictive evaluation, and online policy effects, and clarify how they support the method's contribution.

AuthorsOpenReview ↗

Response to Question 1: Learning Objective, Risk Modeling, and Online Decisions

Question: What substantive advances does DRSR introduce over the closest methods in its learning objective, risk mechanism, or online algorithm? Is the contribution methodological or primarily a systematic integration?

Thank you very much for framing the novelty question at these three concrete levels. This distinction helped us articulate the methodological contribution more precisely and connect each design choice to its experimental evidence.

Our contribution includes specific learning and risk-based decision mechanisms for agent-history pruning, together with their systematic integration around the same joint deletion action. The central advance is to align the supervision unit, risk representation, and online decision with the actual joint deletion to be executed.

AspectTypical singleton-scoring/aggregation approachDRSR
Learning objectiveScore individual Blocks, then rank or aggregate scoresSupervise actual joint deletions using their harm H to the recorded output
Risk modelingScore Blocks independently or model only within-set relationsCondition on the pre-deletion state, deleted and retained information; predict signed joint harm and unsafe probability
Online decisionsCompress using importance, budgets, or fixed thresholdsCombine predicted joint risk with protocol validity, Recent2 protection, cardinality and request budgets; retain the request if no set passes the gate

1. Learning objective: supervise joint deletion directly

Offline, DRSR jointly deletes protocol-valid history Blocks and teacher-forces the same recorded next output under the original and modified contexts. The NLL change defines joint harm H.

This target corresponds directly to the deletion action considered at deployment. The interaction analysis in Response 1 shows reconstruction errors even when exact singleton harms are available, with greater error and risk underestimation as deletion cardinality increases. This provides a concrete modeling rationale for set-level supervision.

2. Risk mechanism: jointly represent the deleted set, retained context, and current state

DRSR predicts the risk of a particular deletion action in the current state. Its representation includes the deleted set, retained information, and pairwise interactions to capture redundancy, complementarity, and dependence among historical information.

The retained-set and pair-pooling ablations test these components' effects in online decisions. The external prediction comparison separately evaluates risk discrimination, ranking, and numerical regression against learned singleton aggregation.

3. Online decisions: risk-constrained selection with abstention

At deployment, predicted joint risk is combined with tool-protocol validity, Recent2 protection, deletion cardinality, and the request-level token budget. A valid set passing the risk gate can be deleted; otherwise the original request is retained.

Risk prediction therefore influences both which deletion is executed and whether pruning occurs. The changes in deletion intensity and Reward after removing abstention provide complementary evidence for this decision mechanism.

We will make the contribution explicit in the revision:

  1. Joint deletion-harm supervision: supervise protocol-valid joint deletions through their effect on the same recorded next output.
  2. State-conditioned set-context risk modeling: jointly represent the current state, deleted set, retained information, and interactions.
  3. Risk-constrained selective deletion: connect predicted joint risk to online selection through protocol constraints, budgets, and abstention.

Once again, we sincerely thank you for recognizing the practical value of this work and its interaction-modeling motivation, and for encouraging us to clarify the technical contribution. Your suggestions helped us organize the method's positioning, mechanisms, and evidence more clearly. We deeply appreciate the time and effort you have devoted to improving this work, and wish you continued success and a fulfilling, enjoyable academic exchange at this year's ICLR.

Comments on the paper

1 comment

We have substantially strengthened the paper in response to the reviewers' comments. Specifically, we further clarified the methodological contributions of joint deletion-harm supervision, state-conditioned set-level risk modeling, and risk-constrained pruning with abstention. We added held-out risk-prediction evaluations, downstream reward validation, cross-extractor experiments, and sensitivity analyses of the harm threshold and risk-control parameters. We also introduced stronger comparisons against random deletion, learned singleton scoring, and exact-H oracle policies, together with more comprehensive baseline implementation details, rollout-cost decomposition, and execution-log analyses.

We sincerely thank all three reviewers for their careful reading, rigorous assessment, and constructive suggestions. We are equally grateful to the Area Chairs and Senior Area Chairs for their time and guidance. Their feedback has been instrumental in helping us clarify the methodological contributions, strengthen the experimental validation, and improve the overall presentation of the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.