Pith.

Artificial Intelligence · arXiv:2609.34649 v1 · submitted 28 Sep 2026 · 65 references

Beyond Skill Evolution: Self-Evolving Context Management Policies for Long-Horizon Agent Harnesses

Weiyuan Li, Jinghan Xu, Aili Chen, Xintao Wang, Shuang Liang, Jiaqing Liang, Deqing Yang

The paper says

“Harness evolution improves LLM agents by learning from execution trajectories, but existing experience- and skill-based methods are less effective on long-horizon tasks.”

more of the abstract

As interactions grow, useful evidence can be buried by redundant or outdated context, making context management itself a key bottleneck. We introduce ContextEvo, a framework that learns a context policy from long-horizon trajectories. ContextEvo reconstructs the model-visible context at key decision points, identifies context-related failures, and applies targeted policy updates. Starting from the open-source Pi-agent harness, ContextEvo improves performance across three long-horizon task benchmarks, achieving results comparable to or better than several prominent agent harnesses, including Codex, OpenCode, and OpenClaw. Additional analyses show that fixed or locally evolved context strategies can fall short under long-horizon information pressure, while our methods adapt to the information demands of each environment.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 29 Sep 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. A genuinely novel abstraction and an unusually honest paper, but the empirical claim rests on single-rollout scores without error bars and full-set gains that include the held-in split, so the improvement is not yet protected from noise. This line is wrong

4 major objections · 4 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 65 references say.

The paper says, at Appendix A.2 / Table 2

The review did not quote this passage.

Pith says: Objection 1 of 4 · major

Appendix A.2 states that each task is attempted once, and no seed, error-bar, or significance information accompanies any score in Table 2. All reported deltas are therefore single-rollout point estimates, and in task counts several headline gains are small: the LHTB full-set gain of 3.1 pp over 46 tasks is one to two tasks, and the DeepSWE iter-2 gain of 7.1 pp over 113 tasks is about eight tasks against a binomial standard error of roughly 4.5 pp.

more of objection 1

The central claims (b102-b107) are not yet separated from run-to-run noise; the paper should report multiple seeds or per-split variance/significance estimates, or otherwise bound the noise floor.

The paper says, at Section 5.1, Table 2 (Held-out column) / Section 7

The review did not quote this passage.

Pith says: Objection 2 of 4 · major

The full-benchmark (All) improvements that carry the headline claims include the held-in split on which the policy was evolved, so they do not separate a real context-policy effect from fitting to the evolution set. The appropriate evidence, the held-out column, is mixed: at iter-1 under GPT-5.6-Luna, DeepSWE falls from 18.5% to 15.4% and BrowseComp-Plus from 39.8% to 33.7%, and under DeepSeek-V4-Flash BrowseComp-Plus falls from 82.7% to 81.6%.

more of objection 2

The paper's conclusion acknowledges that 'held-out changes vary by setting,' but given the one-attempt protocol, the load-bearing premise that CONTEXTEVO improves performance on held-out tasks beyond the noise floor is not yet established and needs a dedicated treatment.

The paper says, at Section 6.4, Table 5

The review did not quote this passage.

Pith says: Objection 3 of 4 · major

The leave-one-operator-out ablation, which is the main support for the claim that each pipeline stage contributes, is itself reported on single runs with no variance, and the scores swing widely (held-out drops from 43.12% for Full to 26.19% without Paginated Trajectory Review and to 35.29% without Counterfactual Replay).

more of objection 3

On held-out splits of 26 tasks these differences may be within the single-rollout noise floor, so the ablation needs the same error-bar/significance treatment as the main results, and the paper should state why the ablation is restricted to LHTB rather than applied to at least one other benchmark.

The paper says, at Sections 4.2 and 7 (b140, b403)

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.76):

“These forms of evidence do not by themselves establish a task-level causal reward contribution for an individual policy mechanism. 9 AIUSE STATEMENT In this work, we used generative AI tools to assist with result interpretation, literature organization, and the drafting, editing, and formatting of text, tables, and figures.”

Pith says: Objection 4 of 4 · major

The method's feedback loop rests on LLM-judge failure attribution, and the authors correctly disclose that the evidence forms do not establish a task-level causal reward contribution for an individual mechanism. What remains open is not the disclosure but the consequence: because all downstream gains depend on that attribution, a controlled test of attribution validity — for example verifying replay predictions …

more of objection 4

… against full-task rerolls on a small held-out subset, or an isolated-mechanism ablation that isolates one policy dimension at a time — would substantially raise the evidential value of the main claims. As written, the paper cannot distinguish a genuinely better context policy from a reward-neutral policy that happens to pass the judge's screens.

4 minor objections

The paper says, at Sections 5.2 and B.2

Pith says: minor objection 1 of 4

The text refers to 'Table 5.1' in two places when the experiments table is numbered Table 2; these cross-references should be corrected.

The paper says, at Appendix C.4, Table 12

Pith says: minor objection 2 of 4

The cross-harness validation uses Qwen3.8-27B as the judge while the main runs use the fixed base model as the judge; the paper should state whether the judge-model choice could confound the cross-harness comparison, since the judge participates in the attribution stage.

show 2 more

The paper says, at Section 6.1, Table 3

Pith says: minor objection 3 of 4

The token figures in Table 3 should state whether they are provider billing tokens or the audit estimate described in Table 7; the two measures can differ substantially and Table 7's note suggests the reported numbers are estimates.

The paper says, at Figure 3

Pith says: minor objection 4 of 4

The bar-panel labels (e.g., '1/20, 16/20') are not mapped to the multi-select context-problem categories in the figure itself; adding a legend or category axis would make the prevalence counts interpretable without going to the appendix.

What would settle it

Re-run all six benchmark-model settings in Table 2 with at least five independent evaluation seeds for the Pi-agent baseline and for the frozen iter-1 and iter-2 policies on the same held-in/held-out splits, and compute bootstrap 95% confidence intervals on the held-out and full deltas. If any advertised improvement has a held-out CI that includes zero, or if the mean held-out delta across settings is no larger than the bootstrap noise floor, the central claim should be reported as a point-estimate comparison and qualified by held-out-only results.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 29 Sep 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic partly agrees 1 min 49 s · 26,113 tokens read, 12,392 written
  3. Editor wrote the verdict above 19 s · 27,659 tokens read, 940 written
  4. Referee major revision: 4 major objections, 4 minor 31 s · 26,311 tokens read, 5,145 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

You should know two things about arXiv:2609.34649. First, the central framing is real: context management as an evolvable, full-lifecycle policy is a genuine bottleneck in long-horizon harnesses, and the three-dimensional abstraction (input assembly, history maintenance, orchestration) is a new contribution beyond the ReSum/ACON/TACO compression line and beyond skill evolution. The evidence-reconstruction, failure-attribution, counterfactual-replay, constrained-update loop is coherent and concretely instantiated through the eight intervention points and the ContextProgram adapter. Second, the paper is unusually honest. It discloses full prompts and protocols, reports held-out separately, and repeatedly states that the evidence does not establish task-level causal reward for individual mechanisms (b140, b403, b406, b460, b473, b491). That is more than most papers in this space do, and it deserves credit.\n\nThe stress-test note is correct, and I agree with the conditional verdict. Appendix A.2 states each task is attempted once, with no seeds and no error bars anywhere in Table 2. In task counts the gains are thin: on LHTB's 46 tasks, 3.1pp is roughly one to two tasks; on DeepSWE's 113, the 7.1pp iter-2 gain is about eight tasks against a binomial standard error near 4.5pp. Held-out results are mixed and sometimes negative: DeepSWE under Luna falls from 18.5 to 15.4 at iter-1, and BrowseComp-Plus falls from 39.8 to 33.7 under Luna and from 82.7 to 81.6 under DeepSeek at iter-1. Because the headline full-set scores include the held-in split on which the policy was evolved, the consistent iter-2-over-iter-1 gains do not yet separate a real context-policy effect from evaluation noise and held-in fitting. The load-bearing premise is the LLM-judge failure attribution: every downstream gain rests on it, and there is no controlled isolated-mechanism ablation to verify it. The authors acknowledge this squarely, so it is an unresolved evidential gap, not a hidden flaw. The component ablation is single-benchmark and single-run as well; the -12.15pp from removing paginated review is suggestive but not decisive.\n\nThis paper deserves a serious referee, not a desk reject. The concept is sound, the engineering is real, and the reporting practice is a model. But the empirical claim is underpowered as written. The fix is clear and addressable: multiple seeds with error bars, and held-out-only deltas given headline prominence. Accept conditionally on those. For anyone working on agent harnesses or self-evolving context management, this is worth reading and worth citing for the abstraction and the evaluation discipline.

The strongest objection

Appendix A.2 states that each task is attempted once, and no seed or error-bar information accompanies any score in Table 2. All reported deltas are therefore single-rollout point estimates. In task-count terms several are small: on LHTB's 46 tasks a 3.1 pp full-set gain is about one to two tasks; on DeepSWE's 113 tasks the 7.1 pp iter-2 gain is about eight tasks against a binomial standard error of roughly 4.5 pp. Held-out results are mixed and sometimes negative: at iter-1 under GPT-5.6-Luna, DeepSWE falls from 18.5% to 15.4% and BrowseComp-Plus from 39.8% to 33.7%, and under DeepSeek-V4-Flash BCP held-out falls from 82.7% to 81.6%. Because the full-benchmark scores include the held-in split on which the policy was evolved, the consistent full-set improvements (b105) do not yet separate a real context-policy effect from evaluation noise and held-in fitting. The authors disclose the one-attempt protocol and state that held-out changes vary by setting (Section 7), and they acknowledge the attribution limit (b140), so this is an unresolved evidential gap rather than a hidden flaw. The load-bearing empirical premise, that ContextEvo improves long-horizon performance beyond the noise floor and on held-out tasks, is not yet established.

The reader's CONDITIONAL verdict is well calibrated, and the paper deserves credit for releasing splits, prompts, and per-task results (Appendix C.3), which makes the missing piece evaluative replication rather than reimplementation. I do not single out failure-attribution validity as the weakest premise, because the authors explicitly and repeatedly acknowledge that their evidence does not establish task-level causal reward for individual mechanisms (b140, b403, b406), and the component ablations in Table 5 give at least some single-run support for the pipeline stages. The more directly load-bearing premise for the paper's headline is statistical: every score is a single rollout (A.2), no confidence intervals are reported, and several held-out deltas are negative while full-set gains are consistently positive. Since the full set includes the held-in tasks used for evolution, the full-set improvements do not yet demonstrate generalizable gains. A multi-seed replication with held-out-only reporting would settle whether the concern lands; until then the conditional verdict stands unchanged.

The referee's report

If the result holds, the paper makes a timely and interesting contribution: it repositioning harness evolution from skill/rule extraction to full-lifecycle context management, and it does so with a carefully staged evidence pipeline (raw/annotation/stratified/bootstrap/local/prospective evidence bases) and an unusually honest treatment of what each evidence type can and cannot support. The token-efficiency result (0.448 reward at 7.52M tokens vs 0.417 at 8.57M for the baseline) is a valuable secondary claim, and the external held-out evaluation, transparent disclosure of the one-attempt protocol, and explicit attribution limits (b140, b403) are genuine strengths. However, the load-bearing empirical premise — that CONTEXTEVO improves long-horizon performance beyond the noise floor and on held-out tasks — is not yet established, because every score in Table 2 is a single-rollout point estimate with no seeds or error bars, and the held-out deltas are mixed and sometimes negative.

Pith's reading of the paper (a machine's summary, not the authors')

This paper argues that on long-horizon tasks the main bottleneck for LLM agents is not the skills they accumulate but how the surrounding harness manages the context the model actually sees — what information enters, what is retained or compacted, and what is shared between subagents. ContextEvo treats that context management as an evolvable policy spanning three dimensions — Input Assembly, History Maintenance, and Context Orchestration — and updates it from execution trajectories while keeping the base model fixed. Starting from the Pi-agent harness, the evolved policies improve full-benchmark scores by 3.1 to 4.4 percentage points under DeepSeek-V4-Flash across three long-horizon benchmarks, and second-round updates improve over first-round ones in all six benchmark–model combinations. The sharpest contrast is that skill-only evolution on the same tasks degrades performance (reward falls from 0.417 to 0.370 while tokens rise from 8.57M to 9.61M), whereas the context-policy evolution reaches 0.448 with 7.52M tokens — evidence that on long horizons the information life cycle, not the skill bank, is the load-bearing lever.

What carries the argument. The central object is the context policy $P = (P^{\mathrm{in}}, P^{\mathrm{hist}}, P^{\mathrm{orch}})$, which maps all information that has entered the system, $I_{\le t}$, to the model-visible context $c_t = P(I_{\le t})$ from which the base model $M$ samples its actions. The evolution loop that carries the argument has three stages: trajectory evidence reconstruction (a full-trajectory information-flow scan plus a paginated review that reconstructs the model-visible context at decision points and carries a running note across pages), failure attribution (an analysis agent that compares successful and failed trajectories and uses counterfactual replay — editing only the context at a fixed decision point and re-sampling the next action — when static evidence is inconclusive), and constrained policy update (an editor that modifies, combines, or introduces mechanisms along the identified dimensions under safety constraints). The counterfactual replay, $a_t^{(r)} \sim M(\cdot|c_t)$ versus $\tilde{a}_t^{(r)} \sim M(\cdot|\tilde{c}_t)$, tests the local sensitivity of the next action to a proposed context change.

Core claim. ContextEvo's central claim is that context management is itself the proper target of harness evolution for long-horizon agent tasks. Rather than extracting reusable skills or rules from trajectories, the framework reconstructs the model-visible context at key decision points, attributes failures to one of three context-policy dimensions (input assembly, history maintenance, context orchestration), and applies constrained updates. The evolved policies adapt to each environment's information pressure: selective observation handling and persistent execution state for terminal workflows, specification and verification continuity for repository-engineering tasks, and recoverable evidence organization for multi-hop search. The paper reports that the method beats the Pi-agent baseline on the full task sets of LHTB (45.3%), DeepSWE (71.7%), and BrowseComp-Plus (88.5%) under DeepSeek-V4-Flash, ranks second only to the Luna-native Codex harness on two benchmarks under GPT-5.6-Luna, and that the second evolution iteration improves over the first in all six benchmark–model settings. Full-policy evolution also outperforms static compression (ReSum, ACON), local compression evolution (TACO), and skill-only evolution on the same tasks.

The premise it rests on. Every downstream gain rests on the failure-attribution stage correctly identifying which of the three context-policy dimensions caused a failure, a judgment the authors themselves concede is supported only locally: their counterfactual replay measures next-action sensitivity rather than task-level causal reward, and they do not verify the attribution with controlled isolated-mechanism ablations.

If the paper is right (Pith's inferences, not the authors')

Skill-evolution methods can actively hurt long-horizon performance: on the same 46 LHTB tasks, skill-only evolution dropped reward from 0.417 to 0.370 while raising mean token use from 8.57M to 9.61M, whereas ContextEvo reached 0.448 with 7.52M tokens.

Fixed or locally evolved context strategies (ReSum, ACON, TACO) all fall short of full-lifecycle policy evolution on the tested benchmarks, so context management should be evolved across the whole information life cycle rather than compression alone.

A learned context policy can transfer across base models within the same environment (the LHTB policy improves both DeepSeek-V4-Flash and GPT-5.6-Luna), but not reliably across environments, where renewed evolution is required.

The method is not tied to one harness: applying ContextEvo to OpenCode and OpenClaw on BrowseComp-Plus hard174 improved OpenCode from 35.06% to 40.80% and OpenClaw from 18.97% to 45.40%.

Of the three evolution stages, paginated trajectory review is the most consequential: removing it on LHTB drops the held-out score from 43.12% to 26.19%.

Because the authors concede that their counterfactual replay measures only local next-action sensitivity, the honest reading is that the gains may come substantially from trajectory-review-driven policy tuning rather than replay-confirmed causality; a controlled study isolating replay's contribution would clarify which stage carries the improvement.

The observed skill failures — invocation gaps, local goal drift, and partial workflow coverage — suggest a division of labour: context policies keep task evidence available, while skills remain usable only when the context that should trigger them is properly maintained, so context-policy evolution may be a prerequisite for skill evolution to pay off on long horizons.

If context-policy evolution generalizes beyond these three benchmarks, it reframes the cost of agent self-improvement: instead of growing skill banks, agents would grow the machinery that decides what enters, persists, and crosses scopes — a change that is token-cheaper and plausibly easier to verify.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Harness evolution improves LLM agents by learning from execution trajectories, but existing experience- and skill-based methods are less effective on long-horizon tasks. As interactions grow, useful evidence can be buried by redundant or outdated context, making context management itself a key bottleneck. We introduce ContextEvo, a framework that learns a context policy from long-horizon trajectories. ContextEvo reconstructs the model-visible context at key decision points, identifies context-related failures, and applies targeted policy updates. Starting from the open-source Pi-agent harness, ContextEvo improves performance across three long-horizon task benchmarks, achieving results comparable to or better than several prominent agent harnesses, including Codex, OpenCode, and OpenClaw. Additional analyses show that fixed or locally evolved context strategies can fall short under long-horizon information pressure, while our methods adapt to the information demands of each environment.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-09-29

The authors' main claim, as the classifier picked it out 0.77

“Starting from the open-source Pi-agent harness, ContextEvo improves performance across three long-horizon task benchmarks, achieving results comparable to or better than several prominent agent harnesses, including Codex, OpenCode, and OpenClaw.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.08 · states limitation 0.21 · human participants 0.04 · controlled comparison 0.76 · randomized assignment 0.02 · states quantitative result 0.09 · compares with prior results 0.97 · mentions code or data availability 0.10

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Harness evolution improves LLM agents by learning from execution trajectories, but existing experience- and skill-based methods are less effective on long-horizon tasks. background 0.71 0.09 0.07 asserted 1.00
As interactions grow, useful evidence can be buried by redundant or outdated context, making context management itself a key bottleneck. background 0.76 0.10 0.07 asserted 0.94
We introduce ContextEvo, a framework that learns a context policy from long-horizon trajectories. contribution 0.96 0.92 0.12 asserted 1.00
ContextEvo reconstructs the model-visible context at key decision points, identifies context-related failures, and applies targeted policy updates. method 0.99 0.92 0.10 asserted 0.99
Starting from the open-source Pi-agent harness, ContextEvo improves performance across three long-horizon task benchmarks, achieving results comparable to or better than several prominent agent harnesses, including Codex, OpenCode, and OpenClaw. result 1.00 0.95 0.54 asserted 1.00
Additional analyses show that fixed or locally evolved context strategies can fall short under long-horizon information pressure, while our methods adapt to the information demands of each environment. result 0.64 0.88 0.28 asserted 0.99

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.97

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-29: 1 request, 4,952 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (6)
Figure 1 of the paper
Figure 1. From skill evolution to context-policy evolution. As interaction histories grow, skill evolution may fail at both stages: extracting low-quality skills during evolution and invoking them at the wrong time or following them unreliably during evaluation. In contrast, our CONTEXTEVO evolves an efficient context policy, achieving higher scores with fewer tokens. 1 arXiv:2609.34649v1 [cs.AI] 28 Sep 2026
Figure 2 of the paper
Figure 2. The overview of CONTEXTEVO. The context policy governs how information is pre￾sented, organized, and shared through the agent harness while the base model remains fixed. CON￾TEXTEVO reconstructs trajectory evidence, attributes context-management failures to policy dimen￾sions, and applies constrained updates for the next evolution round. 4.1 TRAJECTORY EVIDENCE RECONSTRUCTION Evidence Reconstruction establishes the evidence for subsequent failure attribution by recovering how information is exposed and carried through each trajectory. It combines a trajectory-wise information-flow…
Figure 3 of the paper
Figure 3. Environment-specific context problems and policy transfer. Bars show the share of audited cases with each multi-select context problem; labels report counts. The heatmap shows signed score changes (percentage points) from applying the LHTB-evolved policy under DeepSeek and Luna
Figure 4 of the paper
Figure 4. Benchmark coverage. The left panel reports the full inventory and the held-in/held-out split. The sunburst shows the benchmark inventory in the inner ring and task-family composition in the outer ring. LHTB shows its seven largest categories plus an “other” bucket, DeepSWE shows request types, and BrowseComp-Plus is a uniform multi-hop deep-research setting. A BENCHMARK AND EXPERIMENTAL PROTOCOL A.1 BENCHMARK SETTINGS Benchmark profiles. We evaluate three task settings with different sources of long-horizon con￾text. Long-Horizon-Terminal-Bench (LHTB) contains terminal workflows…
Figure 5 of the paper
Figure 5. Long-horizon workload profile. The three panels show the distributions of tool calls, agent turns, and estimated trajectory text volume in frozen Pi/DeepSeek-V4-Flash baseline traces. The token measure is an audit estimate from serialized analyst-visible text rather than provider billing tokens; the corresponding median and interquartile values are listed in
Figure 6 of the paper
Figure 6. Repair-feature inventory by implementation status. Bars count deduplicated repair clusters in the audited policy artifacts, grouped by response feature and benchmark. “Loaded/eval￾uated” means the feature is present in a configured evaluated bundle; it does not establish that the feature fired or caused a reward change. Artifact-only and proposal-only items are shown separately. This inventory complements the case-prevalence figure in Section 6.3; the generic bulk/repeated￾output problem label is not counted as a repair feature. source fragments, JSON reports, and tables. Thus, the…

Figures are the authors', reproduced from the paper.

References (65)
  1. [1] Zhou, Shuyan and Xu, Frank F. and Zhu, Hao and Zhou, Xuhui and Lo, Robert and Sridhar, Abishek and Cheng, Xianyi and Ou, Tianyue and Bisk, Yonatan and Fried, Daniel and Alon, Uri and Neubig, Graham , booktitle =. 2024 , url = on Pith
  2. [2] Liu, Xiao and Yu, Hao and Zhang, Hanchen and Xu, Yifan and Lei, Xuanyu and Lai, Hanyu and Gu, Yu and Ding, Hangliang and Men, Kaiwen and Yang, Kejuan and Zhang, Shudan and Deng, Xiang and Zeng, Aohan and Du, Zhengxiao and Zhang, Chenhui and Shen, Sheng and Zhang, Tianjun and S... on Pith
  3. [3] Jimenez, Carlos E. and Yang, John and Wettig, Alexander and Yao, Shunyu and Pei, Kexin and Press, Ofir and Narasimhan, Karthik , booktitle =. 2024 , url = on Pith
  4. [4] Bai, Yushi and Tu, Shangqing and Zhang, Jiajie and Peng, Hao and Wang, Xiaozhi and Lv, Xin and Cao, Shulin and Xu, Jiazheng and Hou, Lei and Dong, Yuxiao and Tang, Jie and Li, Juanzi , booktitle =. 2025 , doi = on Pith
  5. [5] Yuan, Mengqi and Zhou, Zilong and Xiong, Xinzhuang and Wu, Weiming and Sun, Jiayang and Song, Jiamin and Cui, Kaiqian and Wang, Bowen and Wu, Haoyuan and Li, Yitong and Lu, Dunjie and Lu, Haikong and Zhen, Qi and Wang, Xinyuan and Deng, Jiaqi and Yang, Yuhao and Chen, Cheng an... on Pith
  6. [6] Wei, Jason and Sun, Zhiqing and Papay, Spencer and McKinney, Scott and Han, Jeffrey and Fulford, Isa and Chung, Hyung Won and Passos, Alex Tachard and Fedus, William and Glaese, Amelia , journal =. 2025 , url = on Pith
  7. [7] Chen, Zijian and Ma, Xueguang and Zhuang, Shengyao and Nie, Ping and Zou, Kai and Sharifymoghaddam, Sahel and Liu, Andrew and Green, Joshua and Patel, Kshama and Meng, Ruoxi and Su, Mingyi and Li, Yanxi and Hong, Haoran and Shi, Xinyu and Liu, Xuye and Oyarhoseini, Hosna and T... on Pith
  8. [8] Mialon, Gr. The Twelfth International Conference on Learning Representations , year = on Pith
  9. [9] Huang, Wenqi and Lee, Charley and Tng, Leonard and Ge, Serena , journal =. 2026 , url = on Pith
  10. [10] Lost in the Middle: How Language Models Use Long Contexts , author =. Transactions of the Association for Computational Linguistics , volume =. 2024 , publisher =. doi:10.1162/tacl_a_00638 , url = on Pithdoi:10.1162/tacl_a_00638
  11. [11] Toward Reliable Context Compression for Long-Horizon Agents: An Empirical Study of Execution Instability , author =. arXiv preprint arXiv:2608.06503 , year = on PithPith's reviewarXiv
  12. [12] Jiang, Huiqiang and Wu, Qianhui and Lin, Chin-Yew and Yang, Yuqing and Qiu, Lili , booktitle =. 2023 , doi = on Pith
  13. [13] Packer, Charles and Wooders, Sarah and Lin, Kevin and Fang, Vivian and Patil, Shishir G. and Stoica, Ion and Gonzalez, Joseph E. , journal =. 2023 , url = on Pith
  14. [14] Ye, Rui and Zhang, Zhongwang and Li, Kuan and Yin, Huifeng and Tao, Zhengwei and Zhao, Yida and Su, Liangcai and Zhang, Liwen and Qiao, Zile and Wang, Xinyu and Xie, Pengjun and Huang, Fei and Zhou, Jingren and Chen, Siheng and Jiang, Yong , booktitle =. 2026 , url = on Pith
  15. [15] A Self-Evolving Framework for Efficient Terminal Agents via Observational Context Compression , author =. arXiv preprint arXiv:2604.19572 , year = on PithPith's reviewarXiv
  16. [16] Yao, Yilun and Huang, Shan and Dai, Elsie and Tan, Zhewen and Duan, Zhenyu and Jia, Shousheng and Jiang, Yanbing and Yang, Tong , booktitle =. 2026 , doi = on Pith
  17. [17] Li, Xiaochuan and Ming, Ryan and Chu, Meng and Shao, Shuai and Jin, Rong and Xiong, Chenyan , journal =. 2026 , url = on Pith
  18. [18] Li, Mo and Xu, L. H. and Tan, Qitai and Ma, Long and Song, Hongyong and Cao, Ting and Liu, Yunxin , booktitle =. Sculptor: Empowering. 2026 , url = on Pith
  19. [19] Context as an Environment: Programmatic Context Management for Long-Horizon Agents , author =. arXiv preprint arXiv:2608.21690 , year = on PithPith's reviewarXiv
  20. [20] Alzubi, Salaheddin and Provenzano, Noah and Bingham, Jaydon and Chen, Weiyuan and Vu, Tu , journal =. 2026 , url = on Pith
show 45 more
  1. [21] Zhang, Hanrong and Fan, Shicheng and Zou, Henry Peng and Chen, Yankai and Wang, Zhenting and Zhou, Jiayu and Li, Chengze and Huang, Wei-Chieh and Yao, Yifei and Zheng, Kening and Liu, Xue and Li, Xiaoxiao and Yu, Philip S. , journal =. 2026 , note = on Pith
  2. [22] Yang, Yifan and Gong, Ziyang and Huang, Weiquan and Yang, Qihao and Zhou, Ziwei and Huang, Zisu and Li, Yan and Gao, Xuemei and Dai, Qi and Liu, Bei and Qiu, Kai and Yang, Yuqing and Chen, Dongdong and Yang, Xue and Luo, Chong , journal =. 2026 , url = on Pith
  3. [23] Agrawal, Lakshya A. and Tan, Shangyin and Soylu, Dilara and Ziems, Noah and Khare, Rishi and Opsahl-Ong, Krista and Singhvi, Arnav and Shandilya, Herumb and Ryan, Michael J. and Jiang, Meng and Potts, Christopher and Sen, Koushik and Dimakis, Alexandros G. and Stoica, Ion and ... on Pith
  4. [24] Yang, Zherui and Liu, Fan and Ning, Yansong and Liu, Hao , booktitle =. 2026 , publisher =. doi:10.1145/3770855.3818002 , url = on PitharXiv
  5. [25] Prompting Best Practices: Long-Horizon Reasoning and State Tracking , year = on Pith
  6. [26] DeepPlanning: Benchmarking Long-Horizon Agentic Planning with Verifiable Constraints , author =. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =. 2026 , doi = on Pith
  7. [27] Long-Horizon-Terminal-Bench: Testing the Limits of Agents on Long-Horizon Terminal Tasks with Dense Reward-Based Grading , author =. arXiv preprint arXiv:2607.08964 , year = on PithPith's reviewarXiv
  8. [28] The Horizon Gap: Planning, Memory, Execution, Training, and Evaluation for Long-Horizon LLM Agents , author =. arXiv preprint arXiv:2608.06663 , year = on PithPith's reviewarXiv
  9. [29] Reinforcement Learning for Long-Horizon Interactive LLM Agents , author =. arXiv preprint arXiv:2502.01600 , year = on PithPith's reviewarXiv
  10. [30] AgentGym-RL: An Open-Source Framework to Train LLM Agents for Long-Horizon Decision Making via Multi-Turn RL , author =. International Conference on Learning Representations , volume =. 2026 , url = on Pith
  11. [31] Peng, Jiangweizhi and Liu, Yuanxin and Zhou, Ruida and Fleming, Charles and Wang, Zhaoran and Garcia, Alfredo and Hong, Mingyi , journal =. 2026 , url = on Pith
  12. [32] Demystifying Reinforcement Learning for Long-Horizon Tool-Using Agents: A Comprehensive Recipe , author =. arXiv preprint arXiv:2603.21972 , year = on PitharXiv
  13. [33] Plan-and-Act: Improving Planning of Agents for Long-Horizon Tasks , author =. Proceedings of the 42nd International Conference on Machine Learning , volume =. 2025 , url = on Pith
  14. [34] Language Agent Tree Search Unifies Reasoning, Acting, and Planning in Language Models , author =. Proceedings of the 41st International Conference on Machine Learning , volume =. 2024 , url = on Pith
  15. [35] Qi, Qiuyi and Liang, Tian and Bao, Mutian and Zhang, Jinjian and Liu, Dongnan and Zhou, Wei and Mo, Linjian and Kong, Ming and Liu, Jie and Zhang, Feng and Zhu, Qiang , booktitle =. 2026 , doi = on Pith
  16. [36] A Goal Without a Plan Is Just a Wish: Efficient and Effective Global Planner Training for Long-Horizon Agent Tasks , author =. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =. 2026 , doi = on Pith
  17. [37] Hu, Mengkang and Chen, Tianxing and Chen, Qiguang and Mu, Yao and Shao, Wenqi and Luo, Ping , booktitle =. 2025 , doi = on Pith
  18. [38] Wan, Guangya and Ling, Mingyang and Ren, Xiaoqi and Han, Rujun and Li, Sheng and Zhang, Zizhao , booktitle =. 2026 , doi = on Pith
  19. [39] Kang, Minki and Chen, Wei-Ning and Han, Dongge and Inan, Huseyin A. and Wutschitz, Lukas and Chen, Yanzhi and Sim, Robert and Rajmohan, Saravan , booktitle =. 2026 , note = on Pith
  20. [40] Zhou, Zijian and Qu, Ao and Wu, Zhaoxuan and Kim, Sunghwan and Prakash, Alok and Rus, Daniela and Low, Bryan Kian Hsiang and Liang, Paul Pu , booktitle =. 2026 , url = on Pith
  21. [41] ReSum: Unlocking Long-Horizon Search Intelligence via Context Summarization , author =. arXiv preprint arXiv:2509.13313 , year = on PitharXiv
  22. [42] Li, Ruoran and Zhang, Xinghua and Yu, Haiyang and Duan, Shitong and Li, Xiang and Xiang, Wenxin and Liao, Chonghua and Guo, Xudong and Li, Yongbin and Suo, Jinli , booktitle =. 2026 , doi = on Pith
  23. [43] Memory as Action: Autonomous Context Curation for Long-Horizon Agentic Tasks , author =. Findings of the Association for Computational Linguistics: ACL 2026 , pages =. 2026 , doi = on Pith
  24. [44] Agentic Memory: Learning Unified Long-Term and Short-Term Memory Management for Large Language Model Agents , author =. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =. 2026 , doi = on Pith
  25. [45] Beyond the Context Window: Scaling Agentic RL via End-to-end Optimized Context Compression , author =. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =. 2026 , doi = on Pith
  26. [46] Learning Agent-Compatible Context Management for Long-Horizon Tasks , author =. arXiv preprint arXiv:2605.30785 , year = on PithPith's reviewarXiv
  27. [47] Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents , author =. arXiv preprint arXiv:2607.08716 , year = on PithPith's reviewarXiv
  28. [48] Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models , author =. International Conference on Learning Representations , volume =. 2026 , url = on Pith
  29. [49] MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents , author =. arXiv preprint arXiv:2602.02474 , year = on PithPith's reviewarXiv
  30. [50] Self-Harness: Harnesses That Improve Themselves , author =. arXiv preprint arXiv:2606.09498 , year = on PithPith's reviewarXiv
  31. [51] Ma, Ziyu and Huang, Hailang and Zou, Shun and Wang, Yong and Yang, Shidong and Hu, Yiming and Wei, Fei and Chu, XiangXiang , journal =. 2026 , url = on Pith
  32. [52] Grow the Harness, Not the Context: From Strategy-Free Scaffolds to Reusable Specialist Agents , author =. arXiv preprint arXiv:2609.26760 , year = on PithPith's reviewarXiv
  33. [53] Meta-Harness: End-to-End Optimization of Model Harnesses , author =. arXiv preprint arXiv:2603.28052 , year = on PithPith's reviewarXiv
  34. [54] Agentic Harness Engineering: Observability-Driven Automatic Evolution of Coding-Agent Harnesses , author =. arXiv preprint arXiv:2604.25850 , year = on PithPith's reviewarXiv
  35. [55] Evo-Harness: Context-to-Harness Skill Compilation for Self-Evolving Agents , author =. arXiv preprint arXiv:2608.15071 , year = on PithPith's reviewarXiv
  36. [56] HarnessEvolve: Learning from Reference Trajectories for Reliable Agent Self-Evolution , author =. arXiv preprint arXiv:2609.00829 , year = on PithPith's reviewarXiv
  37. [57] Verify Smarter, Evolve Further: Efficient Harness Evolution through Behavior-Aware Verification , author =. arXiv preprint arXiv:2608.27311 , year = on PithPith's reviewarXiv
  38. [58] AgentRx: Diagnosing AI Agent Failures from Execution Trajectories , author =. arXiv preprint arXiv:2602.02475 , year = on PithPith's reviewarXiv
  39. [59] Where LLM Agents Fail and How They Can Learn From Failures , author =. arXiv preprint arXiv:2509.25370 , year = on PitharXiv
  40. [60] Qi, Yunjia and Yin, Zehua and Shi, Xintong and Peng, Hao and Lu, Songyuanyi and Liu, Yixian and Xuan, Richeng and Liu, Yuhong and Hu, Zhichao and Wang, Xiaozhi and Hou, Lei and Xu, Bin and Li, Juanzi , journal =. 2026 , url = on Pith
  41. [61] Causal Agent Replay: Counterfactual Attribution for LLM-Agent Failures , author =. arXiv preprint arXiv:2606.08275 , year = on PithPith's reviewarXiv
  42. [62] When Continual Learning Moves to Memory: A Study of Experience Reuse in LLM Agents , author =. arXiv preprint arXiv:2604.27003 , year = on PithPith's reviewarXiv
  43. [63] Retrieval-Augmented LLM Agents: Learning to Learn from Experience , author =. arXiv preprint arXiv:2603.18272 , year = on PithPith's reviewarXiv
  44. [64] LongHorizon-Harness: Advancing Long-Horizon Agents for Real-World Tasks , author=. arXiv preprint arXiv:2608.01964 , year= on PithPith's reviewarXiv
  45. [65] Beyond Semantic Organization: Memory as Execution State Management for Long-Horizon Agents , author=. arXiv preprint arXiv:2606.06090 , year= on PithPith's reviewarXiv

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment