Pith.

Artificial Intelligence · arXiv:2607.23809 v1 · submitted 26 Jul 2026 · 14 references

ACM: Agentic Context Management for Long Horizon Tasks

Xiaochuan Li, Ryan Ming, Meng Chu, Shuai Shao, Rong Jin, Chenyan Xiong

The paper says

“Agentic tasks are inherently long-horizon and multi-turn, constantly accumulating context through interactions with the environment. Existing context compression methods inevitably incur information loss and are triggered by rigid heuristic rules, leaving them misaligned with the agent's evolving reasoning focus.”

more of the abstract

We propose Agentic Context Management (ACM), a framework that equips agents with purpose-built context editing tools for lossless context management. Inspired by the interaction between short-term and long-term human memory, the agent autonomously decides when to compress its context, offloads discarded content to an external memory system, and queries it on demand for later retrieval. Building on this framework, we further develop a post-training pipeline that constructs high-quality demonstrations of context management and improves model performance on both agentic search and coding tasks. Further analysis reveals that effective context management reduces peak token pressure, enables extended explorations, and yields more consistent solutions across independent trials. Code, data, and model checkpoints are available at https://github.com/lixiaochuan2020/agentic-context-management.

from the abstract

Pith says

Machine review · grok-4.5 · 30 Jul 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. Solid agent-systems paper: agent-timed lossless compression plus a dual-constraint training recipe; headline gains are real but partly entangled with gold-conditioned teacher actions and uneven baseline training. This line is wrong

3 major objections · 6 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 14 references say.

The paper says, at §4.1, §4.2, App. B.5 (correction teacher prompt)

The review did not quote this passage.

Pith says: Objection 1 of 3 · major

The content filters in §4.2 police only the teacher's `think` text (forbidden coach-words, no gold-answer mentions, no unretrieved docids), but the leakage channel runs through action selection and phrasing, which the filters do not cover. App. B.5 explicitly instructs: search_query 'SHOULD use distinctive terms drawn from the gold evidence that the agent hasn't tried'; the commit branch (highest global priority, Step 1) fires when 'in-context evidence supports the gold answer' and the teacher writes `exact_answer` with {correct_answer} in context.

more of objection 1

Combined with rejection sampling that retains only failed trajectories, the supervised replacement actions are systematically answer-directed. This is most load-bearing for BrowseComp-Plus, where train (680) and eval (150) come from the same benchmark and the 27% headline gain lives. Requested fix: an ablation training with (a) injection-only

The paper says, at Table 2, §5.2

The review did not quote this passage.

Pith says: Objection 2 of 3 · major

No uncertainty quantification is given anywhere. The BCP eval is n=150; the base-ACM vs ACM-Post-Trained gap (0.635 vs 0.727) is likely significant, but the DeepSearchQA gap (0.405 vs 0.425) is ~2 points on an unstated n and may be within noise, as may the SWE-Bench gap (0.508 vs 0.530, and 0.489 ReAct vs 0.530).

more of objection 2

Please report per-cell trial counts (is Table 2 Pass@1 averaged over the 4 trials of Fig. 5?), binomial/bootstrap CIs, and the DeepSearchQA eval-set size. This matters because the out-of-domain transfer result is currently the main evidence against the leakage concern in comment 1, so its statistical robustness is load-bearing.

The paper says, at §3, Table 1 ('Lossless' claim)

The review did not quote this passage.

Pith says: Objection 3 of 3 · major

'Lossless' is an architectural property (raw messages persist on disk), but operationally retrieval is mediated by a querier LLM with a budgeted, filtered output (App. B.4), run by the same 9B model — so recovery of needed facts is lossy in practice.

more of objection 3

The claim is central to the framing (Table 1 marks ACM as the only lossless method). Please scope the claim and, ideally, quantify it: e.g., fidelity of query_memory retrieval on known-fact probes, frequency of failed/empty retrievals, or an oracle-retrieval upper bound. The Fig. 6 case study shows query_memory being used productively once; a systematic measurement would be more convincing.

6 minor objections

The paper says, at §6.2

Pith says: minor objection 1 of 6

Typo: 'DeepResearchQA' should be 'DeepSearchQA'. Also, the +Both regression on DeepSearchQA (0.413 < 0.425 for +ACM alone) is noted but not explained; a sentence on why GPT-5.5 distillation hurts here would help.

The paper says, at Table 2

Pith says: minor objection 2 of 6

Formatting glitches: '0.40588.742K' and '0.508 77.646K' are missing column spaces; capitalization of K/k in peak-token entries is inconsistent (63k vs 79K).

show 4 more

The paper says, at Fig. 3 and Fig. 5

Pith says: minor objection 3 of 6

Fig. 3's legend entries ('Timing of MC', population means) are hard to parse without the caption; consider annotating one sawtooth directly. Fig. 5 uses 'pass4' for the consistency metric while §6.1 defines Pass^K with K=4; unify notation and state in the caption that pass@1 is the trial average.

The paper says, at §5.1 / App. B.6

Pith says: minor objection 4 of 6

Answers are graded by an LLM judge (simple-evals template). Please report judge–human agreement or the judge model used, since Pass@1 differences of a few points could be sensitive to grader behavior, especially for DeepSearchQA set-answer questions.

The paper says, at §5.2

Pith says: minor objection 5 of 6

The claim that ACM-Post-Trained 'nearly match[es] open-source models 40× larger' understates it: 0.727 exceeds Qwen3.5-397B-A17B's 0.653 on BCP. Either claim the stronger result (with appropriate caveats about contamination/eval differences) or rephrase.

The paper says, at App. D

Pith says: minor objection 6 of 6

The 4B negative result is valuable. Consider citing prior observations of small-model agentic collapse (e.g., tool-use RL at small scales) and clarifying whether the 4B model saw the same system prompt — the hallucinated 'I can't do real searches' constraint suggests a prompt/template interaction worth ruling out.

What would settle it

Answer-blind teacher ablation, reusing the released artifacts: keep the student rollouts H+/H− fixed, but re-run Phase 2 annotation with the teacher seeing only the trajectory and a binary failure flag — no A*, no gold_docs_block, and the "gold evidence" rule struck from the search_query instructions. Retrain with identical compute and re-evaluate on the BrowseComp-Plus 150-example eval and DeepSearchQA. If the answer-blind student retains most of the gain (BCP within ~5 absolute points of 0.727, DSQA roughly unchanged), the leakage channel is not load-bearing; if BCP drops substantially toward the ACM-base 0.635 while DSQA holds, the 27% headline figure is materially inflated by answer-derived supervision. A cheap auxiliary check: measure n-gram/entity overlap between teacher-proposed search queries in the released training data and the gold answers.

from the skeptic's reading

How the review went

4 readings by 2 models · 30 Jul 2026

  1. Reader holds up, with conditions confidence high
  2. Skeptic agrees 57 s · 24,365 tokens read, 2,980 written
  3. Editor wrote the verdict above 17 s · 25,758 tokens read, 592 written
  4. Referee major revision: 3 major objections, 6 minor 2 min 25 s · 24,822 tokens read, 3,171 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

Punchline: this is a clean systems-and-training package for long-horizon agents—two tools (manage_context / query_memory) that offload raw history and let the agent pull it back, plus a dual-constraint teacher pipeline that teaches when to compress and when not to. The behavioral story holds up: sawtooth context curves, lower peak tokens, more tool calls, and better Pass4 consistency. Open code, data, and checkpoints are a real plus.

What is actually new is less the memory metaphor (MemGPT, AgentFold, ReSum/ACON sit nearby) and more the combination: agent-initiated compression that keeps raw traces, explicit query_memory, and the inject-on-H− / refine-on-H+ annotation loop with on-policy soft distillation. Table 2’s framework-only ACM base already beats ReAct without post-training; DeepSearchQA is OOD and still moves. Fig. 3–6 and the Pass@K analysis are the right kind of evidence for this claim class.

Soft spots, in proportion. The stress-test on Appendix B.5 is fair: the correction teacher is told to craft search queries with distinctive terms from gold evidence, and commit decisions use gold support. §4.2’s filters catch coach-words and answer text, not action choice. That can inflate the BrowseComp-Plus 27% relative gain, especially with train/eval from the same benchmark. It does not sink the paper—OOD and untrained-ACM gains still point to a real compression-timing skill—but the largest number is not cleanly isolated. Second, post-trained ACM is compared to untrained ReSum/ACON/ACE; the “agent-initiated beats fixed threshold” claim among trained systems is thin (base ACM 0.635 vs ~0.61, n=150, no CIs). “Lossless” is also operationally model-mediated via a querier LLM. Novelty is moderate; significance is subfield-real, not theory-shifting.

Who it’s for: people building long-horizon search/coding agents and anyone training context-management policies. Math is light (standard distillation loss); citations are honest about AgentFold. I’d send it to peer review. Engage the work, use the artifacts, and pressure the leakage and trained-baseline gaps in review rather than dismissing the core idea.

The strongest objection

The central empirical claim rests on the dual-constraint post-training pipeline (§4.1), and the paper's own leakage defenses (§4.2 content filters) only police the teacher's `think` text and forbid citing unretrieved docids. But Appendix B.5's correction teacher is told, under "# RULES FOR search_query": "SHOULD use distinctive terms drawn from the gold evidence that the agent hasn't tried." Combined with rejection sampling that retains only failed trajectories, this means the supervised replacement actions are systematically answer-directed recoveries: when the student stalls, the training data teaches it to issue queries phrased with gold-evidence terminology, or to commit at the earliest point where "in-context evidence supports the gold answer" (Step 1 of the teacher's algorithm). A lexical filter on coach-words and answer mentions cannot catch this — the leakage is in the *choice and phrasing of actions*, not in explicit answer text. This is most load-bearing for BrowseComp-Plus, where training (680) and eval (150) are drawn from the same benchmark, so gold-evidence-derived query styles can transfer directly to eval items of the same construction; the 27% relative gain there is the paper's headline number. Two mitigations keep this from being disqualifying: DeepSearchQA is genuinely out-of-domain (no training data) and still shows a 16% relative gain, suggesting real transfer of the compression-timing skill; and the framework-only comparison (ACM base 0.635 vs ReAct 0.570) involves no teacher at all. However, a secondary soft spot compounds the first: Table 2 compares the *post-trained* ACM model against *untrained* ReSum/ACON/ACE baselines, so the "agent-initiated beats fixed-threshold" component of the claim is only tested among untrained systems (0.635 vs 0.608/0.614, n=150, no CIs reported) — a gap within noise. No trained summary-agent baseline exists to isolate the contribution of agent-timed compression from the contribution of post-training itself.

The reader's weakest_assumption identified the right load-bearing region — teacher annotations conditioned on A* — and my stress test sharpens rather than displaces it: the leakage is not merely a residual risk after content filtering, it is explicitly invited by the correction-teacher prompt's instruction to draw query terms from the gold evidence, a channel the §4.2 filters structurally cannot catch. This validates the reader's CONDITIONAL verdict (contingent on leakage audits and baseline parity) rather than overturning it, so I recommend UNCHANGED. I down-weight the other candidate concerns: (1) the "lossless" framing is genuinely model-mediated (summarizer/querier are the same 9B), but the case study in Fig. 6 shows query_memory being exercised productively, and the reader already prices this into medium correctness risk; (2) small eval n (150) with no CIs weakens the *untrained* comparisons (ACM base vs ReSum/ACON differ by ~4 correct answers) but the trained-model gaps (0.727 vs 0.570) are large enough to survive plausible binomial noise; (3) the OOD DeepSearchQA gain and the consistent Pass^4 consistency improvements across epochs (Fig. 5) provide independent behavioral support that the mechanism does real work. The answer-blind retraining test is decisive, cheap (the pipeline and rollouts are open-sourced), and directly maps onto the condition the reader already attached to acceptance.

The referee's report

If the results hold, this is a useful contribution to the growing context-management literature: the agent-initiated framing is clean, the comparison table against ReSum/ACON/ACE/Mem1/SUPO/AgentFold is fair about regime differences, and the release of the full data-generation pipeline, prompts (App. B), training data, and checkpoints is a genuine reproducibility asset that prior work (e.g., AgentFold) withheld. The framework-only ablation (ACM base 0.635 vs ReAct 0.570 on BCP, Table 2) and the out-of-domain DeepSearchQA gain provide evidence that some of the improvement is a transferable skill rather than benchmark-specific fitting. The Pass@K/Pass^K analysis (Fig. 5) and the honest negative result on 4B models (App. D) strengthen the empirical picture.

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

Long agent tasks pile up tool outputs, failed attempts, and reasoning until the working window is full of noise. Most compression schemes cut or summarize on a fixed schedule and discard the original text, so the agent cannot get details back and the timing rarely matches what it is actually thinking about. This paper argues that the agent should treat context management as an explicit action: it compresses when it chooses, parks the raw history in external memory, and pulls pieces back on demand. A teacher–student post-training recipe teaches when to compress and when not to. On search and coding benchmarks, a 9B model with this skill explores longer, uses fewer peak tokens, and answers more consistently across repeated trials than plain ReAct or forced-summary baselines.

What carries the argument. Agentic Context Management (ACM): short-term working context plus long-term external store, operated by manage_context and query_memory, internalized via a dual-constraint teacher pipeline that both inserts missing compressions and replaces premature ones before on-policy distillation.

Core claim. Giving an agent two tools—manage_context (summarize and offload raw turns) and query_memory (retrieve from those archives)—and training it with dual teacher constraints on when to use or refuse them yields lossless, agent-initiated context control that raises Pass@1 on BrowseComp-Plus, DeepSearchQA, and SWE-Bench Verified while cutting peak token load and lifting consistency across independent runs.

The premise it rests on. Teacher labels that see the golden answer, even after filters against leaking it, still teach transferable timing of compression rather than answer-tied shortcuts—and the same-size model used as summarizer and querier recovers enough of the archived trace for “lossless” to hold in practice.

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

Peak context pressure drops because compression happens before the window saturates, not only at a hard threshold.

Agents can run more exploratory tool calls and more turns before hitting the limit.

Correct solutions become more stable across independent trials (Pass@1 and Pass4 rise together).

Smaller open models can approach much larger systems on hard multi-hop search when context skill is trained in.

Context management is a learnable policy, not only an external monitor or fixed summarizer.

If timing of compression is the scarce skill, synthetic dual-constraint data may matter more than ever-larger context windows for multi-hour agents.

The same offload-and-query pattern could transfer to multi-agent handoffs where one agent’s working buffer must stay short while shared disk holds the full trail.

Failure modes of small models that quit in two turns suggest ACM only pays off once the base policy already sustains long rollouts.

Open-sourcing the generation pipeline lowers the barrier to testing whether other tool-use domains (browsers, IDEs) need the same dual constraints.

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 grok-4.5, 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

Agentic tasks are inherently long-horizon and multi-turn, constantly accumulating context through interactions with the environment. Existing context compression methods inevitably incur information loss and are triggered by rigid heuristic rules, leaving them misaligned with the agent's evolving reasoning focus. We propose Agentic Context Management (ACM), a framework that equips agents with purpose-built context editing tools for lossless context management. Inspired by the interaction between short-term and long-term human memory, the agent autonomously decides when to compress its context, offloads discarded content to an external memory system, and queries it on demand for later retrieval. Building on this framework, we further develop a post-training pipeline that constructs high-quality demonstrations of context management and improves model performance on both agentic search and coding tasks. Further analysis reveals that effective context management reduces peak token pressure, enables extended explorations, and yields more consistent solutions across independent trials. Code, data, and model checkpoints are available at https://github.com/lixiaochuan2020/agentic-context-management.

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 · 2026-09-23

The authors' main claim, as the classifier picked it out 0.96

“Further analysis reveals that effective context management reduces peak token pressure, enables extended explorations, and yields more consistent solutions across independent trials.”

How the main claim is stated: asserted 0.99

What the abstract itself states, as probabilities: claims first 0.06 · states limitation 0.09 · human participants 0.05 · controlled comparison 0.12 · randomized assignment 0.01 · states quantitative result 0.04 · compares with prior results 0.32 · mentions code or data availability 0.99

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
Agentic tasks are inherently long-horizon and multi-turn, constantly accumulating context through interactions with the environment. background 1.00 0.07 0.02 asserted 1.00
Existing context compression methods inevitably incur information loss and are triggered by rigid heuristic rules, leaving them misaligned with the agent's evolving reasoning focus. gap 0.72 0.06 0.03 asserted 1.00
We propose Agentic Context Management (ACM), a framework that equips agents with purpose-built context editing tools for lossless context management. contribution 0.93 0.89 0.17 asserted 1.00
Inspired by the interaction between short-term and long-term human memory, the agent autonomously decides when to compress its context, offloads discarded content to an external memory system, and queries it on demand for later retrieval. method 0.97 0.71 0.09 asserted 0.99
Building on this framework, we further develop a post-training pipeline that constructs high-quality demonstrations of context management and improves model performance on both agentic search and coding tasks. method 0.54 0.95 0.08 asserted 1.00
Further analysis reveals that effective context management reduces peak token pressure, enables extended explorations, and yields more consistent solutions across independent trials. result 0.70 0.93 0.46 asserted 0.99
Code, data, and model checkpoints are available at https://github.com/lixiaochuan2020/agentic-context-management. contribution 0.68 0.30 0.02 asserted 0.85

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.93

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, 2026-09-23: 1 request, 5,594 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 (8)
Figure 1 of the paper
Figure 1. Overview of our ACM Framework. ReAct eventually hits the context limit, while the Summary Agent
Figure 2 of the paper
Figure 2. Overview of the dual-constraint training data generation pipeline. A student agent completes task rollouts
Figure 4 of the paper
Figure 4. Per-tool call frequency on BrowseComp￾Plus for GPT-5.5, Qwen3.5-9B, and our post-trained Qwen3.5-9B under the ACM framework. text limit, driven by the agent’s own reasoning state. 2) The payoff of context management is sub￾stantial: By keeping the context compact, ACM substantially slows context growth while enabling more exploratory turns. As a result, the agent can continue reasoning and interacting for significantly longer before reaching the context limit. This ad￾vantage is particularly important for challenging questions that require extended, multi-step explo￾ration and…
Figure 5 of the paper
Figure 5. Accuracy of Qwen3.5-9B on BrowseComp-Plus under ReAct, ACM, and three epochs of ACM post
Figure 6 of the paper
Figure 6. Trajectory of our ACM-Post-Trained Model on a 5-constraint multi-hop question. The model self-monitors its context, alternates manage_context (▼) with query_memory (♦) probes, and terminates only after confirming the working context is small enough. Without compression the trajectory would cross the 128K base-model limit at turn 47 (dashed line); the base model in fact gives up on this question in all 4 runs. 6 Ablation Study 6.1 Study of Pass@K We examine how agentic context management af￾fects Pass@K (Brown et al., 2024; Snell et al., 2024) and PassK (Yao et al., 2024) with K=4,…
Figure 7 of the paper
Figure 7. Running pivot fraction over relative trajectory progress for ReAct, ACM, and ACM-Post-Trained, across
Figure 8 of the paper
Figure 8. Why the model terminates early. In blue the 4B model correctly self-reports that it has used only 23K of its 131K-token budget and could keep searching; in the very next sentence (red) it hallucinates the constraint “I can’t do real searches” and commits to a guess. The failure is not a context-budget failure—it is a failure to maintain a long-horizon plan—so context-management tools have no opportunity to help. 24

Showing 7 of the paper's 8 figures. Figures are the authors', reproduced from the paper.

References (14)
  1. [1] First-person only ("I notice...", "my searches...", "my working context..."). No third-person self-reference, no naming or alluding to any external party on Pith
  2. [2] Ground the rationale ONLY in observable trajectory signals, e.g.: * the same or near-identical query keywords repeated across multiple search turns * accumulated input tokens / context- window pressure * the last N turns produced no new docids and no new evidence * the retriev... on Pith
  3. [3] Resum: Unlocking long-horizon search in- telligence via context summarization.arXiv preprint arXiv:2509.13313. Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh Jing Hua, Zhou- jun Cheng, Dongchan Shin, Fangyu Lei, Yitao Liu, Yiheng Xu, Shuy... on PitharXiv
  4. [4] Length: between 200 and 1200 characters. Output the JSON object only. No preamble, no closing remarks, no prose outside the JSON. Correction teacher: rewrite an unproductive mc step.A complementary teacher targets fail- ure traces dominated byover-compression: it se- lects an ... on Pith
  5. [5] Qwen3 technical report.arXiv preprint arXiv:2505.09388. John Yang, Carlos E Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan, and Ofir Press. 2024. SWE-agent: Agent-computer interfaces enable automated software engineering. InAdvances in Neural Informat... on PithPith's reviewarXiv
  6. [6] Agentfold: Long-horizon web agents with proactive context management.arXiv preprint arXiv:2510.24699. Hongli Yu, Tinghong Chen, Jiangtao Feng, Jiangjie Chen, Weinan Dai, Qiying Yu, Ya-Qin Zhang, Wei- Ying Ma, Jingjing Liu, Mingxuan Wang, and Hao Zhou. 2025. Memagent: Reshaping... on PitharXiv
  7. [10] STRICTLY FORBIDDEN: * mentioning, naming, or describing any part of the golden answer * naming any docid the agent has NOT retrieved in this trajectory * any of these words:`coach`,`coaches`, `coaching`,`feedback`,`review`,` reviewer`,`reviewed`,`external`,` advised`,`advisor`... on Pith
  8. [12] commit --- at SOME mc step, the in- context evidence is already enough to answer. Replace that mc call with a final -answer turn on Pith
  9. [13] replace_with_search --- at SOME mc step, a specific search query you propose would have been more productive than compressing. The query must NOT duplicate any query the agent has already tried on Pith
  10. [14] replace_with_get_document --- at SOME mc step, a specific docid that was ALREADY visible in a prior search-result snippet should have been fetched. The docid must be in the cumulative retrieved set at history[:replace_id]. Algorithm --- global action priority, earliest step wi... on Pith
  11. [2023] MemGPT: Towards LLMs as operating sys- tems.arXiv preprint arXiv:2310.08560. Jiayi Pan, Xingyao Wang, Graham Neubig, Navdeep Jaitly, Heng Ji, Alane Suhr, and Yizhe Zhang. 2024. Training software engineering agents and verifiers with swe-gym.arXiv preprint arXiv:2412.21139. Joo... on PithPith's reviewarXiv
  12. [2024] Theagentcompany: benchmarking llm agents on consequential real world tasks.arXiv preprint arXiv:2412.14161. Wujiang Xu, Zujie Liang, Kai Mei, Hang Gao, Jun- tao Tan, and Yongfeng Zhang. 2025. A-mem: Agentic memory for llm agents.arXiv preprint arXiv:2502.12110. An Yang, Anfeng... on PithPith's reviewarXiv
  13. [2025] Swe-bench pro: Can ai agents solve long- horizon software engineering tasks?arXiv preprint arXiv:2509.16941. Runnan Fang, Yuan Liang, Xiaobin Wang, Jialong Wu, Shuofei Qiao, Pengjun Xie, Fei Huang, Hua- jun Chen, and Ningyu Zhang. 2025. Memp: Ex- ploring agent procedural memor... on PithPith's reviewarXiv
  14. [2026] Agentic context engineering: Evolving con- texts for self-improving language models.Preprint, arXiv:2510.04618. Zijian Zhou, Ao Qu, Zhaoxuan Wu, Sunghwan Kim, Alok Prakash, Daniela Rus, Jinhua Zhao, Bryan Kian Hsiang Low, and Paul Pu Liang. 2025. Mem1: Learning to synergize me... on PithPith's reviewarXiv
Cited by 15 Pith papers

Reviewed papers on Pith that cite this one. The full citation record

  1. Coding Agent Memory Post-Training II: Unlocking the Memory Potential of Pre-trained File Operations for Long-Horizon Tasks via On-Policy Distillation cs.LG · Nov 2026
  2. CompAct-Bench: A Benchmark for Working-Context Compaction in Long-Horizon Agent Tasks cs.AI · Nov 2026
  3. Octopus: One Brain, Many Branches for Latent-KV Multi-Agent Reinforcement Learning cs.LG · Nov 2026
  4. Prune the Summary, Not the History: Source-Judged Compaction for Long-Horizon Agents cs.CL · Oct 2026
  5. Context Language Models cs.AI · Sep 2026
  6. Coding Agent Memory Post-training: Unlocking the Memory Potential of Pre-trained File Operations for Long-Horizon Tasks via Reinforcement Learning cs.LG · Sep 2026
  7. MoE-Context: Show the Model Only What It Needs cs.AI · Nov 2026
  8. CarryOn: Learning What to Keep in Context for Long-Horizon LLM Agents cs.CL · Nov 2026
  9. Monitor Accuracy Isn't the Answer: A Decision-Value Framework for Evaluating AI Agent Restart Monitors cs.AI · Oct 2026
  10. NavHarness: Towards Lifelong Embodied Navigation cs.RO · Sep 2026
  11. Stable Geometry with Divergent Task Evidence for Efficient Long-Horizon Agent Compression cs.AI · Sep 2026
  12. StateComp: Learning When to Compress History in Long Horizon Agents cs.AI · Sep 2026
  13. Memory Control Signals Emerge Before Action in Long Horizon Agents cs.AI · Sep 2026
  14. DRSR: Learning Set-Level Deletion Risk for Efficient Long-Horizon Agents cs.AI · Sep 2026
  15. Context as an Environment: Programmatic Context Management for Long-Horizon Agents cs.AI · Aug 2026
How the literature uses it

What this paper does with the arXiv papers it cites

Each citing sentence is taken from the citing paper's LaTeX source and typed by Pith's classifier. A label appears at probability 0.8 or more (0.95 for support), a contest at 0.7 or more; the quoted words are the citing authors'. How each label was checked · JSON

What it cites · 16 sentences without a confident label

Uses the method or tool of 4 papers
Uses data or values from 1 paper
Compares with 5 papers
  • We compare ACM against three agent frameworks: (1) ReAct [ref] , the standard reasoning-and-acting agent without any context management; (2) Summary Agent [ref] , which triggers summarization when context usage exceeds a fixed threshold; and (3) Memory Agent [ref] , which accumulates experiences from previous rollouts but does not dynamically manage its intra-trajectory context.

    ReAct: Synergizing Reasoning and Acting in Language Models · Experimental Setup · 0.93

  • We compare ACM against three agent frameworks: (1) ReAct [ref] , the standard reasoning-and-acting agent without any context management; (2) Summary Agent [ref] , which triggers summarization when context usage exceeds a fixed threshold; and (3) Memory Agent [ref] , which accumulates experiences from previous rollouts but does not dynamically manage its intra-trajectory context.

    2509.13313 · Experimental Setup · 0.93

  • We compare ACM against three agent frameworks: (1) ReAct [ref] , the standard reasoning-and-acting agent without any context management; (2) Summary Agent [ref] , which triggers summarization when context usage exceeds a fixed threshold; and (3) Memory Agent [ref] , which accumulates experiences from previous rollouts but does not dynamically manage its intra-trajectory context.

    ACON: Optimizing Context Compression for Long-horizon LLM Agents · Experimental Setup · 0.93

Cites as support 2 papers
  • As discussed in Section [ref] and supported by [ref] , even frontier models struggle to determine the appropriate timing for context management.

    2510.06727 · Training Data Generation · 0.99

  • As discussed in Section [ref] and supported by [ref] , even frontier models struggle to determine the appropriate timing for context management.

    2510.24699 · Training Data Generation · 0.99

Cites as background 15 papers
  • [ref] (SUPO) and [ref] both use reinforcement learning [ref] to teach the agent to summarize or fold context.

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models · Heuristic Context Compression · 0.82

  • [ref] (SUPO) and [ref] both use reinforcement learning [ref] to teach the agent to summarize or fold context.

    2510.11967 · Heuristic Context Compression · 0.82

  • [ref] (SUPO) and [ref] both use reinforcement learning [ref] to teach the agent to summarize or fold context.

    2510.06727 · Heuristic Context Compression · 0.82

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