acceptodds
Under review as a conference paper at ICLR 2027

When Can Agents Forget Their Reasoning? ICLR for Long-Horizon Agent Context Compression

Abstract

Long horizon language model agents continually accumulate reasoning history, increasing context length and inference cost even after earlier decisions have been executed and observed. Unlike static Chain of Thought compression, removing historical reasoning can change future actions and the resulting interaction trajectory. We study when such reasoning can be safely forgotten. We propose Interaction Aware Compression for Long Horizon Reasoning (ICLR), a training free online method that ranks reasoning blocks using frozen proxy entropy while preserving actions, tool calls, and observations. Across four domains of long horizon agent tasks, ICLR improves average reward from 0.699 to 0.718, while reducing input, output, and cache read tokens by 25.5%, 14.4%, and 33.3%, respectively. Ablations reveal trajectory amplification, where local reasoning deletion produces nonlinear changes in total computation by altering subsequent interaction. Representation probing, activation patching, and controlled trajectory analyses further suggest that historical reasoning becomes more replaceable once task relevant derived state has been reliably externalized into code, files, tool outputs, or environmental feedback. These results characterize agent reasoning as dynamic working state rather than permanent interaction history.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM Agents, Long Horizon Agents, Context Compression, Reasoning Compression, Reasoning History, Agent Memory, Tool Using Agents, Mechanistic Analysis

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_tT66OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Interaction Aware Compression for Long Horizon Reasoning (ICLR), a training-free technique designed to trim past reasoning tokens in long-horizon LLM agents. Using a frozen secondary model to evaluate the predictive entropy of reasoning blocks, the approach discards the lowest-entropy blocks following each interaction while keeping actions, tool calls, and environment observations intact. Evaluated on WorkBuddyBench with DeepSeek-V4-Flash, ICLR achieves higher task performance alongside a noticeable reduction in memory. Additionally, the authors analyze how pruning reasoning impacts future interaction and examine why historical reasoning becomes redundant once key information is captured externally.

Critical strengths
  • Entropy-driven context compression. While Step Entropy is not a new concept, and it is already used for early stopping of reasoning, this paper extends it to compressing historical reasoning in long-horizon agentic tasks. To my knowledge, this is the first entropy-driven methods for context compression, providing a fresh perspective. The method is simple, training-free, and achieves promising efficiency gains while maintaining competitive task performance.

  • Insightful analysis of reasoning compression in agent interactions. Beyond end-to-end evaluations, the paper provides extensive analyses through trajectory-level ablations, representation probing, activation patching, and controlled state interventions. These experiments offer valuable insights into how reasoning deletion affects future agent behavior and why historical reasoning becomes less necessary once task-relevant information is externalized.

Critical weaknesses
  • Limited evaluation across benchmarks and models. The experiments are restricted to a single benchmark (WorkBuddyBench) and a single acting model (DeepSeek-V4-Flash). Although the benchmark covers four domains, evaluations on additional long-horizon agent benchmarks and different model families would provide stronger evidence of generalizability.

  • Limited analysis of entropy scorer robustness. The method relies on a single frozen Qwen3.5-9B model to estimate reasoning-block entropy. Since entropy estimates can vary across models, evaluating different scorer sizes and model families would help establish whether the proposed compression strategy is robust to the choice of scorer.

Questions for authors

I am more than happy to increase my score weak accept/clear accept if the following questions could be addressed.

  • Generalization: How effectively does ICLR perform when applied to alternative agent models and different long-horizon benchmarks?

  • Scorer Robustness: How sensitive is the framework to the model chosen as the entropy scorer? Could you try other scorers?

4 comments
AuthorsOpenReview ↗

Cross-Benchmark and Cross-Actor Evaluation

Dear Reviewer,

We sincerely thank you for the valuable time and effort devoted to this careful, professional review. We especially appreciate your recognition of entropy-driven historical reasoning compression, the training-free design, and the trajectory ablations, representation probes, activation patching, and externalization analysis. Your positive assessment is very encouraging.

Your suggestions on additional benchmarks, acting models, and Scorer robustness provided clear guidance for examining the method's scope. We report results from three external benchmarks, three acting models, and four Scorers below.

W1: Evaluation across benchmarks and acting models is limited.

The experiments are restricted to a single benchmark (WorkBuddyBench) and a single acting model (DeepSeek-V4-Flash). Although the benchmark covers four domains, evaluations on additional long-horizon agent benchmarks and different model families would provide stronger evidence of generalizability.

Thank you for identifying this limitation. We report runs on τ-bench, Terminal-Bench 2.0, and SWE-bench Verified, together with additional acting models. Evaluation batches are reported separately.

1. Cross-benchmark evaluation

Tokens are per-task averages; k denotes 1,000 tokens.

BenchmarkNBaseICLRΔRewardBase Actor TokensICLR Actor TokensScorer Tokens
τ-bench165113/165 (0.6848)107/165 (0.6485)−0.036497.7k87.2k26.0k
Terminal-Bench 2.08954/89 (0.6067)58/89 (0.6517)+0.04491,010.5k866.0k232.4k
SWE-bench Verified10084/100 (0.8400)93/100 (0.9300)+0.09001,364.7k825.8k424.5k

Both Base and ICLR use the native TauICLRAgent on τ-bench.

These results extend evaluation beyond WorkBuddyBench, but benefits vary: Terminal-Bench and SWE-bench show higher Reward and lower Actor token usage; τ-bench shows lower Actor token usage alongside lower Reward. We report quality and resource usage separately rather than claim improvements on every benchmark.

Local Scorer workload

BenchmarkValid NScorer forwards/taskScorer tokens/taskCUDA interval/task
τ-bench16510.7326,0436.96 s
Terminal-Bench 2.08914.88232,44822.91 s
SWE-bench Verified10023.08424,48341.81 s

Terminal has one recovery/acceptance-layer timing record, and SWE has four recovery timing records; these are listed separately as recovery-process records. Overall Runner timings are reported using each benchmark's valid task count, separately from Scorer CUDA timing.

2. Cross-actor evaluation

We additionally evaluate Qwen3.8-Flash, HY-3, and GPT-5.6-Luna on Full260.

Acting modelNBase RewardICLR RewardBase Actor TokensICLR Actor TokensICLR Scorer Tokens
Qwen3.8-Flash2600.77280.77951,197.9k1,104.8k238.6k
HY-32600.56940.57501,437.8k1,353.4k247.6k
GPT-5.6-Luna2600.55580.5558669.8k694.6k194.9k

Qwen3.8-Flash and HY-3 show slightly higher mean Reward and lower Actor token usage. GPT-5.6-Luna has unchanged mean Reward but increased Actor token usage.

These results expand the range of acting models while showing that benefits depend on the actor. The revision will report these differences fully and restrict generalization claims to the evaluated models and environments.

AuthorsOpenReview ↗

Scorer Sensitivity and Responses to the Two Questions

W2: Analysis of entropy Scorer robustness is limited.

The method relies on a single frozen Qwen3.5-9B model to estimate reasoning-block entropy. Since entropy estimates can vary across models, evaluating different scorer sizes and model families would help establish whether the proposed compression strategy is robust to the choice of scorer.

Thank you very much for this important suggestion. We compare four entropy Scorers across the Qwen and Llama model families, covering different sizes and checkpoints.

We fix DeepSeek-V4-Flash as the acting model and retain the same Full260 task list, 80% Block deletion rule, and evaluation protocol. This sensitivity analysis uses the same Full260 Base reference; each Scorer configuration's ICLR result is calculated from its corresponding run records, separately from the main three-seed experiment.

Scorer backboneNBase OverallICLR OverallBase Actor Tokens/taskICLR Actor Tokens/taskScorer Tokens/task
Qwen3.5-4B2600.74510.72192,558.2k1,918.5k185.3k
Qwen3.5-9B2600.74510.73722,558.2k2,082.2k191.3k
Llama-3.2-3B2600.74510.71232,558.2k2,152.1k171.9k
Llama-3.1-8B2600.74510.73332,558.2k2,173.8k221.4k

All four configurations reduce Actor token usage, but their Rewards in this sensitivity batch are below the Base reference. Within each family's comparison, the larger Scorer achieves higher Reward alongside increased Actor and Scorer token processing.

These results show that Scorer size and checkpoint affect downstream performance and resource usage. We will explicitly report this sensitivity rather than treat different Scorers as interchangeable without performance loss. We sincerely appreciate your suggestion, which helped us characterize the quality-cost trade-off in Scorer selection more accurately.

Q1: How effectively does ICLR perform on alternative agent models and different long-horizon benchmarks?

Thank you very much for your attention to generalization and your specific suggestions. In our response to W1, we report the additional cross-benchmark and cross-actor evaluations, together with Actor usage and local Scorer overhead. Please see the corresponding tables for the detailed results.

The supplementary results show that ICLR can be applied to different task environments and acting models, with quality and usage benefits varying by setting. The revision will present each result separately and clarify the method's scope. We sincerely appreciate your professional guidance, which helped us expand the evaluation and support our conclusions more fully and accurately.

Q2: How sensitive is the framework to the entropy Scorer model? Could you try other Scorers?

Yes. We compare Qwen3.5-4B, Qwen3.5-9B, Llama-3.2-3B, and Llama-3.1-8B with the acting model, task list, and deletion rule fixed.

These results show that entropy scoring can operate across two model families, but downstream quality and resource usage are sensitive to Scorer choice. In this batch, the larger Scorer within each family achieves higher Reward than the smaller version, with increased token processing. We therefore report these as a sensitivity analysis, explicitly showing quality and cost changes for each configuration and qualifying the Scorer-robustness conclusion.

We sincerely thank you again for recognizing the method design and mechanistic analysis, and for proposing these specific, professional experimental directions. Your suggestions helped us expand the evaluation and better understand the benefits and limitations across task environments, acting models, and Scorer configurations.

We are deeply grateful for the valuable time, professional judgment, and care you have devoted to understanding and improving this work, and wish you a fulfilling, enjoyable, and intellectually stimulating academic exchange at this year's ICLR.

Reviewer_tT66OpenReview ↗

Response to rebuttal

Many thanks to the authors for their detailed rebuttal. I am satisfied with their responses, particularly regarding the following points:

  • Generalization: The authors provide additional evaluations on new benchmarks, including τ\tau-Bench and Terminal-Bench v2.0, as well as experiments with other actor models.
  • Scorer Robustness: The authors conduct further evaluations using diverse entropy scorers to assess the sensitivity of their method to scorer choice.

I am therefore raising my score from 2 to 3 as promised. I am not assigning a 4 because, the work is more a clever adaptation of existing concepts (step entropy + ACON). Good luck!

Thank you very much for your feedback! We greatly appreciate your encouragement and recognition. Wish you all the best!

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

The paper proposes ICLR, a training-free method that compresses an agent’s past reasoning during a long-horizon run. It scores paragraph-level reasoning blocks by mean token entropy with a frozen Qwen3.5-9B model and replaces the lowest 80% with special skip block and leaving other components such as actions, tool calls, and observations intact. The author evaluate their methods On WorkBuddyBench and show average reward rises with fewer tokens compared to baseline. The authors also add probing, activation patching, and small controlled interventions to argue that reasoning is dispensable once its derived state has been externalized into code, files, or tool outputs.

Critical strengths

The question is timely and practically relevant: reasoning tokens dominate agent context cost, and the closed-loop framing, where an edit changes future actions, is a useful distinction from static CoT compression. The design is simple and deployable, since it needs no training and only touches reasoning while leaving the external record intact.

Critical weaknesses

My main concern would be the novelty of this paper. The scoring rule is similar to Step Entropy (ICLR 2026) applied online to agent histories with a different scorer. Agent history compression is well covered by ACON (2025) and similar work. The paper also doesn’t engage with prior reports that stripping thinking blocks from earlier turns hurts agentic performance (e.g., MiniMax-M2, DeepTool). Those results seem to conflict with the claim that reasoning is largely forgettable, so the paper needs to reconcile them.

Another concern is the validity of evaluation. The headline reward gain comes entirely from the Security domain, while Code, Office, and Web all get worse. Deleting all thinking seems to perform about the same while saving far more tokens, and on Pilot40 ICLR uses more tokens than the base, so the efficiency claim is weak against baselines. Also token numbers are inconsistent across the abstract and tables (see questions below), and scorer overhead is ignored.

Questions for authors

The abstract reports input/output/cache reductions of 25.5%/14.4%/33.3% (Full260), while the tables report −17% to −18.5% on total tokens and +6.9% on Pilot40. Could you give one consistent table (same tasks, same metric, input+output+cache) that includes the scorer’s own compute? Does ICLR still save net compute or cost once the scorer is included?

4 comments
AuthorsOpenReview ↗

Method Positioning and Historical Thinking

Dear Reviewer,

We sincerely thank you for your careful and professional review, and for recognizing the practical relevance, closed-loop evaluation, and deployability of our approach. Your comments helped us identify reporting omissions and clarify the contribution. We respond point by point below.

Relationship to Step Entropy, ACON, and prior compression work

We agree that entropy scoring and agent-history compression have precedents. Our focus is how continuous editing of historical thinking propagates through subsequent state transitions to affect tool choices, recovery, repeated reasoning, total computation, and final Reward. Edited histories persist in later requests, enabling trajectory-level evaluation. We will sharpen this positioning and expand the related-work discussion.

Reconciling reports that removing earlier thinking harms performance

We agree with the observations cited from MiniMax-M2 and DeepTool. Our hypothesis is conditional: deleting information available only in thinking can impair later decisions; once that information is reliably stored and accessible in code, files, tool outputs, or environmental state, some thinking may become more replaceable.

We expanded fixed-state continuations to 50 branches per condition:

InterventionConditionValid/planned branchesRule-based score: mean ± SD
Office fixed stateKeep all thinking50/500.8718 ± 0.0734
Office fixed stateDelete all thinking50/500.8132 ± 0.0881
Position interventionKeep all thinking50/500.8731 ± 0.0740
Position interventionKeep only latest thinking50/500.8692 ± 0.0715
Position interventionDelete all thinking50/500.6768 ± 0.0950
Position interventionDelete only latest thinking50/500.7584 ± 0.0850
After writing codeKeep newly added thinking50/500.7591 ± 0.0720
After writing codeRemove newly added thinking50/500.7576 ± 0.0730

Deleting all thinking can reduce scores in these states, whereas removing newly added thinking after code is written yields scores close to retaining it.

We also expanded feedback-visibility interventions to 100 branches per condition, starting from the same task and environmental state:

stderrBranchesTool-choice judgmentSame error recurs
Visible100Edit: 100/10023/100 (23%)
Hidden100Bash: 100/10071/100 (71%)

In this fixed task, visible feedback reduces recurrence of the same error by 48 percentage points, consistent with external feedback supporting tool selection and recovery.

Externalization-gap statistics, representation probes, and activation patching provide complementary mechanistic evidence; their protocols are in the corresponding appendices. These findings concern the conditions under which historical thinking becomes more replaceable.

AuthorsOpenReview ↗

Full260 Re-evaluation, Pilot40, and Scorer Costs

Full260 domain differences and execution-environment review

We sincerely thank you for your professional, careful, and thorough examination of the Security results. Your comments prompted us to recheck the experiment records and helped us identify Docker runtime anomalies in some tasks, as well as insufficient run-integrity checks and reporting. We sincerely apologize for these shortcomings.

Baseline and ICLR used the same Docker environment, task configurations, and evaluation procedure, and the anomalies affected both groups. We applied a common integrity standard to inspect and rerun affected tasks; normally completed runs with low Reward were retained.

The revision uses the restored Full260 three-seed results. All methods share the tasks, model configurations, tool environment, and scoring rules, with 780 rollouts per method. The earlier single-seed records and new three-seed results are reported separately.

The restored repeated runs continue to show lower Actor token usage and similar mean Reward, supporting the paper's core conclusion. These corrections improve execution and statistical reporting; the method design and research question remain unchanged.

DomainNBase RewardICLR RewardBase InputBase OutputBase I+OICLR InputICLR OutputICLR I+OScorer tokensScorer CUDA s
Code800.75210.7326534,646742,2411,276,887501,454865,2161,366,670235,31624.61
Office500.81960.8192592,7431,324,1521,916,895505,5311,286,3131,791,844243,61326.08
Security600.69640.74921,192,7251,582,5852,775,3101,084,6341,163,1322,247,766255,31625.73
Web700.72580.72951,148,2953,146,1434,294,438692,1572,563,1323,255,289233,63124.87
Overall2600.7451 ± 0.14730.7523 ± 0.1370862,8961,695,2772,558,173688,1611,472,0772,160,238241,07325.22

Token usage and scorer CUDA time are per-task averages. Overall means are weighted by domain task counts; ± denotes SD across task-level mean Rewards.

Pilot40 and consistent reporting

Thank you for identifying the Pilot40 token increase and inconsistencies in reporting. We will retain the Pilot40 efficiency regression and clearly state the task set, run batch, token-field definitions, and aggregation procedure for each result. Full260, Pilot40, and the short- and long-context sets will be reported separately.

Scorer-inclusive accounting

Per-task averages from the same Full260 three-seed batch are:

MetricBaseICLRChange
Actor Input862,896688,161−20.25%
Actor Output1,695,2771,472,077−13.17%
Actor Input+Output2,558,1732,160,238−15.56%
Local scorer tokens0241,073—
Local scorer CUDA time025.22 s—

The scorer is a locally deployed, frozen Qwen3.5-9B that scores existing thinking rather than generating new long reasoning. Scoring has a 42K context limit; pruning is skipped above that limit without truncating the Actor's history.

Scorer tokens use Qwen's tokenizer and represent local computation rather than additional Actor API billing. The scorer consumes GPU forward computation and model residency resources. Its mean peak allocated and reserved GPU memory are 21.46 GiB and 29.68 GiB, respectively.

We report Actor API usage and local scorer resources separately: Actor Input+Output decreases by 15.56%, while the auxiliary scorer processes 241,073 tokens and uses 25.22 seconds of active CUDA computation per task. This makes the token savings and auxiliary computational overhead explicit.

AuthorsOpenReview ↗

Ablation80 Selection and Short-/Long-Context Results

Incremental value over deleting all thinking

When initially constructing Ablation80, we prioritized tasks with shorter contexts and faster completion in preliminary runs to manage evaluation costs. We sincerely apologize for omitting this selection criterion from the paper, and we are grateful that your comments helped us identify and correct this reporting omission.

The selection was based on preliminary context length and execution cost; final task scores were not used as a selection criterion. Your comment helped us recognize that this task composition may be relevant to our central research question: whether the value of retaining historical thinking changes with context conditions. We therefore added short- and long-context comparisons, whose results provide further support for this conditional view. Please accept our sincere gratitude and respect for your careful review.

Short-context Ablation80

Each method uses three independent seeds per task: 240 rollouts across 80 tasks, with 20 tasks per domain.

MethodCodeOfficeSecurityWebOverall Reward: mean ± SD
ICLR0.66640.81300.73360.77270.7464 ± 0.0986
Delete All Thinking0.63940.73210.68130.71480.6919 ± 0.1023
Random 80% Blocks0.60120.70630.69510.69820.6752 ± 0.1228

For paired analyses, we first average the three seeds within each task, then bootstrap the paired task-level differences. Full260 and Short80 are analyzed separately:

Paired comparisonTasksΔReward95% paired bootstrap CIWin/tie/loss
Full260: ICLR − Base260+0.0071[−0.0161, +0.0307]136/0/124
Short80: ICLR − Delete All80+0.0545[+0.0330, +0.0760]56/0/24

The short-context paired interval supports a mean Reward advantage over Delete All Thinking on this task set. Random 80% Blocks deletes the same proportion of blocks and is reported separately from the token-level random controls below.

Additional long-context evaluation

We selected a nonoverlapping long-context set from tasks outside Short80 using preliminary context length: 20 Code, 10 Office, 25 Security, and 25 Web tasks.

Base, ICLR, Delete All Thinking, and the token-level random controls use the same actor, inference settings, output budget, tools, initial state, and verifier. Each method has three independent seeds per task, totaling 240 rollouts.

All token values below are millions per task. ± denotes SD across task-level means; overall means are weighted by task counts. The random percentages concern tokens, whereas Short80's Random 80% control concerns blocks.

MetricDomainNBaseICLRDelete AllRandom 5%Random 10%Random 20%Random 40%
RewardCode200.65270.67510.54320.62510.63920.61640.5905
RewardOffice100.70620.73660.60410.71720.66050.65500.6080
RewardSecurity250.68180.70410.57890.67930.66410.64080.6149
RewardWeb250.67250.69020.56480.68540.67120.63630.5984
RewardOverall800.6747 ± 0.11040.6966 ± 0.09500.5687 ± 0.09520.6724 ± 0.10000.6596 ± 0.10820.6351 ± 0.12340.6028 ± 0.1018
InputOverall803.2529 ± 0.82852.2391 ± 0.70703.3810 ± 0.85423.5181 ± 0.88933.4916 ± 0.82993.6220 ± 0.89493.6275 ± 1.0233
OutputOverall800.0870 ± 0.01200.0780 ± 0.00880.1334 ± 0.01960.0984 ± 0.01470.0947 ± 0.01020.0994 ± 0.02140.0919 ± 0.0145
I+OCode203.0772.0953.4163.2793.2113.4763.331
I+OOffice103.2332.0773.3233.6893.7433.4673.693
I+OSecurity253.5172.4583.6693.7293.8563.8384.063
I+OWeb253.4162.4503.5153.7453.5543.9033.697
I+OOverall803.3399 ± 0.83592.3171 ± 0.71153.5144 ± 0.86363.6165 ± 0.89883.5863 ± 0.83753.7214 ± 0.90683.7194 ± 1.0332

On this long-context set, ICLR achieves higher mean Reward and fewer total task tokens than Delete All Thinking. These results motivate evaluating local deletion together with subsequent reasoning, tool interactions, and the complete execution trajectory.

We sincerely thank you again for your professional review and constructive guidance, which helped us identify reporting omissions and clarify the contribution, its conditions of applicability, and the quality-cost trade-off.

Dear Reviewer,

We sincerely appreciate your valuable comments and suggestions, which have been tremendously helpful in guiding us toward further improvements to our manuscript.

After carefully considering your feedback, we have completed the corresponding revisions and uploaded the updated PDF version of our manuscript. We hope these changes provide clearer and more comprehensive explanations of the relevant points while further improving the clarity, accuracy, and overall quality of our work.

If you have any remaining questions or concerns, or any additional comments and suggestions, we would be truly grateful for your feedback and would be more than happy to provide further explanations or clarifications. Your insights are invaluable to us and play an important role in helping us refine our manuscript and strengthen our research.

Once again, we sincerely thank you for the valuable time and effort you have devoted to reviewing our paper, as well as for your thoughtful and constructive comments. We deeply appreciate your dedication and contributions to the academic community.

We also wish you good health, happiness, and continued success in your research and professional endeavors! May your academic journey at ICLR be filled with inspiring discussions, exciting research insights, and meaningful discoveries, bringing you a truly fortunate, fulfilling, and enjoyable experience!

With our sincere gratitude and warmest wishes,

The Authors

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

The paper proposes ICLR, a training-free method that deletes low-entropy blocks from an agent’s recently completed reasoning while preserving actions, tool calls, and observations. A frozen Qwen proxy scores the blocks, and subsequent interactions consume the compressed history. Its contributions are the online pruning method, evidence that deletion can nonlinearly change subsequent computation, and an externalization hypothesis: reasoning becomes more replaceable once useful derived information is stored elsewhere. Validation uses DeepSeek-V4-Flash on 260 WorkBuddyBench tasks across four domains, with a 40-task baseline comparison, 80-task ablations, representation probing, activation patching, and controlled continuations. The full evaluation reports reward increasing from 0.699 to 0.718, with 18.5% fewer input-plus-output tokens.

Critical strengths
  1. The evaluation captures the consequences of compression during actual interaction. Deletions persist into subsequent agent requests, allowing tool use, recovery, and termination to change. This is consequential because the observed nonmonotonic relationship between deletion and total computation demonstrates why evaluating compression on fixed transcripts can miss important system effects.
  2. The externalization hypothesis has useful preliminary support. Within-task comparisons associate unexternalized derived state with a 19.75-percentage-point increase in later reasoning reuse, with a task-bootstrap interval of [1.00, 37.05] points. Controlled continuations also show preserved scores after reasoning is cleared following code creation. These findings provide a potentially useful explanation of reasoning persistence beyond reporting token savings, although their scope remains limited.
Critical weaknesses
  1. The incremental value of entropy-guided pruning remains uncertain. On the 80-task ablation, deleting all reasoning achieves 71.78 versus ICLR’s 73.76 reward, using 1.11M versus 2.49M tokens (Table 3). The full-benchmark gain is driven entirely by Security, while the other three domains regress (Table 2). No uncertainty estimates or repeated-run results accompany these comparisons, and scorer-inclusive cost is unclear. These limitations matter because a much simpler intervention captures most of the reward improvement at substantially lower reported token usage.
Questions for authors
  1. How many independent runs support Tables 2–3, and what paired uncertainty estimates are available for the reward differences?
4 comments
AuthorsOpenReview ↗

Task Selection and Short-Context Ablation

Dear Reviewer,

We sincerely thank you for the valuable time and effort devoted to reviewing our work and for your specific comments on incremental value, domain differences, statistical reliability, and system costs. We especially appreciate your analysis of the Ablation80 controls. Your comments prompted us to recheck task selection and revisit our central question: does the role of historical Thinking change with context length and task difficulty?

Tables 2 and 3 originally used one seed per task; all supplementary repeated experiments below use three independent seeds per task. Thank you for recommending independent repetitions, which reduce dependence on a single trajectory and provide more reliable estimates of mean performance and variability.

1. Task selection and motivation

For the initial Ablation80, we prioritized tasks with shorter contexts and faster completion in preliminary runs to control costs. We omitted this selection criterion from the paper and sincerely apologize for this reporting oversight. We appreciate your comments, which helped us identify and correct it.

We would like to clarify that task selection was motivated solely by computational cost, rather than by an intention to obtain more favorable scores. After reviewing the results in light of your comments, we recognized that this selection criterion may be related to our original research motivation. The results also provide indirect support for this view. Please allow us to express our sincere appreciation and respect once again for your valuable insight.

Difficult tasks typically require longer contexts and more API calls, motivating this cost-based selection. Consequently, this cohort primarily characterizes short-context compression. Such tasks can more readily recover state from current instructions and tool feedback. Long-context tasks may depend more on historical plans, constraints, and intermediate derivations; deleting them may require state reconstruction, increasing repeated reasoning, tool interactions, and generation costs.

Thank you for helping sharpen our question: does the effect of Thinking deletion depend on context length, and can selective retention reduce repeated reasoning and additional computation in long-context tasks? We therefore report repetitions on the short-context 80 tasks and a separate, non-overlapping long-context 80-task cohort.

2. Short-context Ablation80: three-seed repetitions

This cohort contains 80 tasks, 20 per domain. Each method uses three independent seeds per task, totaling 240 rollouts/method.

Reward by domain

Domain (N=20 each)ICLRDelete All ThinkingRandom 80% Blocks
Code0.66640.63940.6012
Office0.81300.73210.7063
Security0.73360.68130.6951
Web0.77270.71480.6982
Overall (N=80)0.74640.69190.6752

Paired comparison with Delete All Thinking

We first average Reward over three seeds within each task, then bootstrap the paired differences across 80 tasks.

MetricResult
Paired tasks80
ΔReward (ICLR − Delete All)+0.0545
95% paired bootstrap CI[+0.0330, +0.0760]
Win / Tie / Loss56 / 0 / 24

The interval excludes zero, supporting ICLR's mean Reward advantage over deleting all Thinking on this cohort. ICLR has higher task-level mean Reward on 56 tasks and lower Reward on 24.

We recognize the cost competitiveness of deleting all Thinking on short-context tasks; the paired analysis supports ICLR's Reward advantage on this cohort. In the independent long-context 80-task experiment, ICLR and Delete All Thinking achieve mean Rewards of 0.6966 and 0.5687, with Actor Input+Output of 2.3171M and 3.5144M tokens/task, respectively. Thus, on that long-context cohort, selective deletion achieves higher performance and lower Actor token usage: more aggressive local deletion does not necessarily reduce total task usage. We report the two context settings separately to clarify benefits and scope.

AuthorsOpenReview ↗

Run Validation, Full260 Results, and Scorer Costs

1. Docker validation and three-seed repetitions

We sincerely appreciate your attention to the Security results. Our checks identified Docker runtime failures only in Security tasks, affecting both Baseline and ICLR. We reran affected tasks under a uniform integrity criterion; normally completed tasks with low Reward remained in the statistics.

Given stochastic generation and multi-turn interaction, we additionally used three seeds per task under the same tasks, model configuration, tool environment, and scoring rules to reduce dependence on individual trajectories. Changes in other domains arise from rerunning tasks and seed-dependent trajectory variation.

The supplementary results continue to support reduced trajectory-level token usage with comparable observed mean performance. These additions improve runtime-integrity checks and repeatability evaluation; the method design and research question remain unchanged.

2. Full260: domain results and paired uncertainty

Thank you for highlighting domain differences. Base and ICLR each use three independent seeds per task, totaling 780 rollouts/method. Each seed starts from the same initial task state, with model version, temperature, reasoning effort, tool environment, task budget, and verifier fixed.

All token and CUDA figures below are per-task averages. I+O denotes Input+Output; Overall is weighted by domain task counts.

DomainNBase RewardICLR RewardBase InputBase OutputBase I+OICLR InputICLR OutputICLR I+OScorer TokensScorer CUDA (s)
Code800.75210.7326534,646742,2411,276,887501,454865,2161,366,670235,31624.61
Office500.81960.8192592,7431,324,1521,916,895505,5311,286,3131,791,844243,61326.08
Security600.69640.74921,192,7251,582,5852,775,3101,084,6341,163,1322,247,766255,31625.73
Web700.72580.72951,148,2953,146,1434,294,438692,1572,563,1323,255,289233,63124.87
Overall2600.74510.7523862,8961,695,2772,558,173688,1611,472,0772,160,238241,07325.22

Actor API usage and local Scorer processing are reported separately. CUDA time measures active local computation, not end-to-end latency or a unified monetary cost.

Full260 paired results

Paired tasksΔReward (ICLR − Base)95% task-bootstrap CIWin / Tie / Loss
260+0.0071[−0.0161, +0.0307]136 / 0 / 124

The Full260 paired confidence interval crosses zero, so we do not claim a statistically significant overall Reward improvement. In this batch, ICLR nevertheless reduces Actor Input+Output by 15.56%, with observed mean Reward of 0.7523 versus Base's 0.7451, demonstrating a trade-off between lower Actor usage and similar observed mean performance. In the added long-context 80-task experiment, ICLR achieves both higher mean Reward and lower Actor token consumption than Base, further supporting selective retention of historical reasoning on that cohort.

3. Scorer cost accounting

The Scorer is a locally deployed, frozen Qwen3.5-9B that scores entropy on existing Thinking without generating additional long reasoning. Its overhead primarily comes from forward computation. Scoring has a 42K-token context limit; when exceeded, pruning is skipped without truncating Actor history.

On Full260, the Scorer processes 241,073 tokens/task on average, with 25.22 seconds of active CUDA computation. It incurs no additional third-party API token bill, but consumes local GPU resources, so we report its computation separately from Actor API usage.

AuthorsOpenReview ↗

Long-Context Ablation and Revision Summary

1. New long-context Ablation80

Using preliminary-run context lengths, we selected a long-context cohort of 80 tasks disjoint from the short-context cohort: Code 20, Office 10, Security 25, and Web 25. More tasks were selected from Web and Security because these domains contained more long-context tasks.

Under the same model, inference parameters, output budget, tool environment, initial state, and verifier, we compare Base, ICLR, Delete All Thinking, and Random 5%, 10%, 20%, and 40% Token deletion. Each method uses three independent seeds per task, totaling 240 rollouts.

These random controls delete by token proportion and are reported separately from Random 80% Blocks in the short-context experiment.

Token figures are in millions/task. Overall entries with ± report the mean and SD across task-level means; domain-weighted aggregation uses the listed task counts.

MetricDomainNBaseICLRDelete AllRandom 5%Random 10%Random 20%Random 40%
RewardCode200.65270.67510.54320.62510.63920.61640.5905
RewardOffice100.70620.73660.60410.71720.66050.65500.6080
RewardSecurity250.68180.70410.57890.67930.66410.64080.6149
RewardWeb250.67250.69020.56480.68540.67120.63630.5984
RewardOverall800.6747 ± 0.11040.6966 ± 0.09500.5687 ± 0.09520.6724 ± 0.10000.6596 ± 0.10820.6351 ± 0.12340.6028 ± 0.1018
InputOverall803.2529 ± 0.82852.2391 ± 0.70703.3810 ± 0.85423.5181 ± 0.88933.4916 ± 0.82993.6220 ± 0.89493.6275 ± 1.0233
OutputOverall800.0870 ± 0.01200.0780 ± 0.00880.1334 ± 0.01960.0984 ± 0.01470.0947 ± 0.01020.0994 ± 0.02140.0919 ± 0.0145
Input+OutputCode203.0772.0953.4163.2793.2113.4763.331
Input+OutputOffice103.2332.0773.3233.6893.7433.4673.693
Input+OutputSecurity253.5172.4583.6693.7293.8563.8384.063
Input+OutputWeb253.4162.4503.5153.7453.5543.9033.697
Input+OutputOverall803.3399 ± 0.83592.3171 ± 0.71153.5144 ± 0.86363.6165 ± 0.89883.5863 ± 0.83753.7214 ± 0.90683.7194 ± 1.0332

2. Revision scope and interpretation

We sincerely thank you again for the valuable time and effort you have devoted to this careful and rigorous review. Your specific, professional suggestions on control methods, domain differences, repeated runs, and system costs helped us identify and address shortcomings in our experimental reporting, and clarify the method's benefits and scope.

The supplementary experiments indicate that the value of selectively retaining historical Thinking should be assessed in relation to context conditions. On the short-context 80 tasks, the paired results support ICLR's Reward advantage over deleting all Thinking. On the new long-context 80 tasks, the mean Reward difference relative to deleting all Thinking is larger, alongside lower Actor token usage, showing more pronounced benefits in both quality and usage. These findings further support retaining some historical reasoning on the evaluated long-context tasks.

The revision will present Full260 and the two purposefully selected 80-task cohorts separately, explicitly documenting task selection, runtime-environment handling, valid task counts, seed counts, paired intervals, and Scorer overhead so readers can assess the results and their scope accurately.

We are deeply grateful for the time, professional judgment, and care you have devoted to understanding and improving this work. Please accept our sincere appreciation and our wishes for a fulfilling, enjoyable, and intellectually stimulating academic exchange at this year's ICLR.

Dear Reviewer,

We sincerely appreciate your valuable comments and suggestions, which have been tremendously helpful in guiding us toward further improvements to our manuscript.

After carefully considering your feedback, we have completed the corresponding revisions and uploaded the updated PDF version of our manuscript. We hope these changes provide clearer and more comprehensive explanations of the relevant points while further improving the clarity, accuracy, and overall quality of our work.

If you have any remaining questions or concerns, or any additional comments and suggestions, we would be truly grateful for your feedback and would be more than happy to provide further explanations or clarifications. Your insights are invaluable to us and play an important role in helping us refine our manuscript and strengthen our research.

Once again, we sincerely thank you for the valuable time and effort you have devoted to reviewing our paper, as well as for your thoughtful and constructive comments. We deeply appreciate your dedication and contributions to the academic community.

We also wish you good health, happiness, and continued success in your research and professional endeavors! May your academic journey at ICLR be filled with inspiring discussions, exciting research insights, and meaningful discoveries, bringing you a truly fortunate, fulfilling, and enjoyable experience!

With our sincere gratitude and warmest wishes,

The Authors

Comments on the paper

1 comment

We sincerely thank all three reviewers for their careful reading, thoughtful assessments, and constructive feedback. We are also deeply grateful to the Area Chair and Senior Area Chairs for their valuable time, dedicated efforts, and guidance throughout the review process.

In response to the reviewers' suggestions, we have strengthened our evaluation through repeated runs, paired statistical analyses, cross-benchmark and cross-model experiments, and more comprehensive reporting of computational overhead. Beyond addressing these concerns, we have also conducted additional systematic comparisons under short- and long-context conditions, extended our sensitivity analysis across different entropy scorers, and performed further controlled interventions to investigate the relationships among historical Thinking, environmental feedback, and subsequent agent behavior. We hope these additional experiments and analyses will help the reviewers and Area Chair gain a clearer and more comprehensive understanding of our research motivation, core contributions, and experimental findings.

Once again, we sincerely appreciate the considerable time, effort, and expertise that all reviewers and chairs have devoted to evaluating our work. Your thoughtful suggestions have not only helped us strengthen the experimental evidence but also inspired valuable directions for further research. We are truly grateful for your patience, professional judgment, and constructive engagement, and we will carefully incorporate your feedback into the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.