Pith.

Artificial Intelligence · arXiv:2609.40169 v1 · submitted 30 Sep 2026 · 49 references

Learning from Research: Toward Lifelong Agent Harness Evolution

Jingbo Yang, Kwei-Herng Lai, Xiaowen Wang, Yaar Harari, Evgeniy Gabrilovich, Shiyu Chang

The paper says

“Language agents are expected to solve increasingly complex tasks, creating a growing need for continual improvement.”

more of the abstract

One promising approach is to evolve the agent harness, the software that governs tool use, memory management, and task execution, while keeping the underlying language model fixed. Recent methods automate this process by using a meta coding agent to modify the harness based on execution feedback. However, relying on that agent's existing knowledge and observed failures can restrict exploration and make adaptation reactive. Inspired by how human experts learn from the research literature for new solutions, we introduce ScholarEvolve, a framework that automatically draws on state-of-the-art research to guide harness evolution. ScholarEvolve organizes the harness evolution directions into functional modules and uses topic modeling to identify distinct improvement strategies for each module. It implements these strategies and evaluates their combinations to improve task performance. Moreover, the framework is designed to incorporate new publications over time, allowing research advances to drive proactive lifelong evolution. Experiments demonstrate improvements on AppWorld and Tau2-Bench. ScholarEvolve raises Qwen3.5-27B task goal completion from 49.6% to 63.6% on AppWorld Challenge, and raises GPT-5.4-mini pass@1 from 72.7% to 81.9% on Tau2-Bench Telecom.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 1 Oct 2026 · what was read was not recorded · not checked by a person

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

ScholarEvolve puts a genuinely new proposal source under harness evolution — the research literature, curated by topic modeling — and backs it with the cleanest held-out evaluation in this line of work; the one real question is whether that single 57-task dev split generalizes. This line is wrong

the referee pass did not run

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 49 references say.

What would settle it

Re-run the full Qwen3.5-27B AppWorld evolution with a re-split validation set: draw a new 57-task D_val' by bootstrap-resampling the existing D_val tasks (or by sampling from the D_evo + D_val pool), keep D_evo for artifact construction, and repeat mutation selection, crossover, and the three generation cycles with the same budget, research pool, and publication windows. Evaluate each re-split champion on D_N_test and D_C_test. If the 12.4-point Normal TGC gain over the initial harness does not replicate within roughly half its magnitude across 5-10 re-splits, the central generalization claim is sensitive to the particular dev draw. A cheaper ancillary check: enlarge the crossover shortlist from the current cap of four combinations to all non-trivial combinations on D_val, and test whether the additive ranking of Eq.

from the skeptic's reading

How the review went

3 readings by deepseek-v4-flash · 1 Oct 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic partly agrees 1 min 14 s · 30,126 tokens read, 10,026 written
  3. Editor wrote the verdict above 20 s · 31,833 tokens read, 796 written

Not run on this paper: the referee, 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

ScholarEvolve is the first credible demonstration that the research literature, curated by topic modeling, can serve as the proposal source for harness evolution, and it evaluates that claim more honestly than most papers in this area. If I worked on agent harnesses, this is what I'd build on.

What's genuinely new: prior work (Meta Harness, AHE, AgentSquare) draws mutation ideas from execution feedback or code search. ScholarEvolve pulls mechanisms from papers, organizes them into orthogonal topics per module, and recombines the resulting candidates. That is a real addition to the search-proposal space, not a tweak. The setup is careful too: D_test is untouched by every selection decision, persistent artifacts are frozen before test evaluation, the Meta Harness baseline is budget-matched, and they report means and standard deviations over three runs.

The strongest evidence is the Challenge transfer. Amazon and Gmail never appear in evolution or validation, and the harness still improves there. Gains of 9 to 14 points on two backbones are believable because the protocol is clean. I also credit the residual failure audit, and the honest disclosure that Qwen's above-training-max subgroup interval includes zero.

Where I'd push back, in proportion: the single 57-task dev split is the load-bearing assumption. Every choice — which mutations stick (Eq. 4), which combinations get shortlisted (Eq. 3), which champion wins across all three lifelong generations — is made on that one split with three evaluation runs. The Challenge transfer mitigates this but does not show that a different dev-split draw would select an equally good harness. That is an addressable gap, and a re-split robustness run would close it.

Two smaller items. The additive crossover score in Eq. 3 assumes independent module gains, but the paper's own Table 3 shows the misranking is caught by joint evaluation, so this is a genuine but minor concern, not a flaw. And the paired-bootstrap retention rule never gives alpha a number, so selection strictness is unspecified. Both are fixable.

One prose issue: comparing the evolved Qwen to Kimi-K2.6 and GPT-5.4 is apples-to-oranges, since those are standalone frontier models, not harness-tuned. That comparison should be descriptive, not head-to-head. It doesn't touch the central claim.

Verdict: solid paper, deserves a serious referee. The conditions are fixable, and the core result — literature-guided proposals work — looks real.

The strongest objection

The reported test numbers are honest measurements: D_test is untouched, artifacts are frozen, the Meta Harness baseline is budget-matched, split APIs (Amazon/Gmail) are absent from evolution and validation, and standard deviations are small. The real risk is upstream of those numbers. Every selection decision — per-module mutation retention via the paired bootstrap of Eq. 4, the additive crossover shortlist of Eq. 3, and the champion choice across all three lifelong generations — is made on a single fixed 57-task dev split with only three evaluation runs. I do not put the primary weight on the additivity assumption in Eq. 3: the paper's own Table 3 (a 73.7% skills candidate beating a 75.4% one once combined, 80.7% -> 84.8%) shows interactions are caught, because joint evaluation (Eq. 5) is the actual selection criterion and the additive score only prunes the shortlist. The under-mitigated concern is representativeness: if the 57 D_val tasks (max 10 reference APIs, specific application mix) are not a faithful draw from P_tar, the harness is tuned to D_val noise, and the reported gains — though real for this draw — flatter the method's expected outcome. This compounds in lifelong evolution, where selection is repeated on the same 57 tasks across three publication windows (Fig. 3, Table 5). The Challenge transfer is the strongest mitigation, but the beyond-training-max subgroup in Table 4 has Qwen's 95% interval include zero ([-0.7, 19.6]), which the paper honestly discloses. Nothing here breaks the central claim; the gap is the absence of any evidence that the same method would select an equally good harness under a different dev-split draw, and no sensitivity analysis of the shortlist size is provided.

The reader's weakest_assumption puts the Eq. 3 additivity first and D_val representativeness second. I partially agree but invert the emphasis. The additivity risk is real yet structurally mitigated: Eq. 3 only builds the shortlist, while the retention rule (Eq. 4) and champion selection operate on jointly evaluated gains (Eq. 5), and Table 3 is direct evidence that a misranked standalone candidate was corrected at combination time. The representativeness of the 57-task D_val is less mitigated because it conditions every decision — including compounding choices across three lifelong generations — and the paper offers no re-split or repeated-search stability evidence. That said, this is an addressable robustness gap, not a flaw in the reported measurements: the test protocol is clean, transfer to unseen applications is substantial, and the gains are large relative to the three-run standard deviations. These strengths support the reader's CONDITIONAL verdict, so I recommend no change to it. I give credit where due: test splits held out from all optimization, frozen persistent artifacts, budget-matched Meta Harness baseline, and the honestly disclosed wide interval on the above-training-maximum API subgroup. The proposed re-split test is the single check that would settle whether the generalization claim survives a different dev-set draw; if it passes across several re-splits, the conditional concerns are fully resolved.

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

This paper tries to establish that a language agent's execution software, the harness, can be improved continuously by learning from the research literature rather than from the agent's own experience and failures alone. The proposed framework, ScholarEvolve, divides the harness into five modules, clusters retrieved papers into distinct mechanism topics per module, implements each as a candidate mutation, and recombines the best candidates before selecting a new champion. On held-out test tasks the evolved Qwen3.5-27B harness improves AppWorld task goal completion from 69.0% to 81.4% on Normal and from 49.6% to 63.6% on Challenge, and the evolved GPT-5.4-mini harness improves Tau2-Bench Telecom pass@1 from 72.7% to 81.9%. The claim that matters is that new publications can drive further gains over simulated time—proactive lifelong evolution around a fixed backbone—and that the gains transfer to applications the agent never saw during development.

What carries the argument. The load-bearing objects are the five-module decomposition of the harness (tool interface, context management, skills, memories, agentic workflows) with typed interfaces that let module implementations be swapped and recombined; topic modeling conditioned on each module's responsibility, which merges overlapping mechanisms into orthogonal families so the implementation budget covers distinct strategies rather than near-duplicates; the mutation blueprint, a paper-derived spec that separates the source mechanism from host-harness adaptation and preserves provenance; and the additive crossover score of Eq. 3, which reuses per-module measurements to shortlist combinations, paired with joint evaluation (Eq. 5) as a regression check and the paired-bootstrap retention rule of Eq. 4. Topic orthogonality is what makes recombination legal: modules with separate responsibilities and mechanisms with separate loci are expected to compose without conflict, an expectation the joint evaluation is meant to verify.

Core claim. Organizing harness search by module and by topic gives broader, more effective exploration than letting a coding agent modify the harness from its own knowledge of observed failures. ScholarEvolve builds a research pool from audits of failed trajectories, clusters papers into semantically orthogonal mechanism families per module, distributes the candidate budget round-robin across topics so distinct mechanisms are covered, and implements paper-derived mutations behind typed interfaces. Candidates are measured as single-module changes, ranked for crossover by additive predicted gain on the validation set, and shortlisted combinations are evaluated jointly to catch interactions, with a new champion retained only when a paired-bootstrap interval on validation gain lies above zero. The authors attribute the consistent wins over both the starting harness and the budget-matched Meta Harness to this design, and show the gains persist across three simulated publication windows on AppWorld Normal, rising from 69.0% to 81.5% task goal completion over four generations.

The premise it rests on. Selection and crossover trust that gains measured on a small validation split (57 tasks on AppWorld) generalize to the held-out test distribution, and that the additive crossover score predicts combined gains; the paper's own residual error and the swapped-skills example in Table 3 show that additivity can mislead, which only the joint evaluation of shortlisted combinations partially corrects.

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

A harness built from research guidance can close most of the capability gap between a modest backbone and a much stronger one: Qwen3.5-27B reduces its task goal completion gap to GPT-5.4 on AppWorld Normal from 16.7 to 4.3 points, about 74% of the gap, with model weights frozen.

Evolved harnesses and their derived artifacts transfer to applications absent from development (Amazon and Gmail) and to tasks requiring more reference APIs than any training task, with positive mean gains beyond the training maximum.

New publications initiate further search cycles from the retained champion, so a deployed agent can harden itself against anticipated failures before those failures accumulate in its own interactions.

Evolution improves reliability as well as coverage: on AppWorld, Qwen's tasks solved in all three evaluation runs rose from 205 to 333, and roughly three quarters of its gains came from tasks that only intermittently succeeded initially.

Because the method separates mechanism from adaptation via blueprints, the same pipeline could plausibly mine subfields such as safety, reward modeling, or tool design without changing the core loop—a testable extension the paper does not claim.

The known failure of additive ranking suggests a cheap upgrade: a second-order interaction term, or interaction-aware shortlisting, could reduce wasted joint evaluations, and the paper's own Table 3 indicates such a term would add signal.

The residual-failure taxonomy (requirement tracking and evidence coverage barely move) implies harness evolution has diminishing returns on those failure families without backbone or training changes, which could guide where future effort should go.

The small 57-task validation set is the practical ceiling on how much search can be trusted; a stress test evolving on a deliberately skewed validation split would reveal how quickly the reported gains degrade on out-of-distribution tasks.

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

Language agents are expected to solve increasingly complex tasks, creating a growing need for continual improvement. One promising approach is to evolve the agent harness, the software that governs tool use, memory management, and task execution, while keeping the underlying language model fixed. Recent methods automate this process by using a meta coding agent to modify the harness based on execution feedback. However, relying on that agent's existing knowledge and observed failures can restrict exploration and make adaptation reactive. Inspired by how human experts learn from the research literature for new solutions, we introduce ScholarEvolve, a framework that automatically draws on state-of-the-art research to guide harness evolution. ScholarEvolve organizes the harness evolution directions into functional modules and uses topic modeling to identify distinct improvement strategies for each module. It implements these strategies and evaluates their combinations to improve task performance. Moreover, the framework is designed to incorporate new publications over time, allowing research advances to drive proactive lifelong evolution. Experiments demonstrate improvements on AppWorld and Tau2-Bench. ScholarEvolve raises Qwen3.5-27B task goal completion from 49.6% to 63.6% on AppWorld Challenge, and raises GPT-5.4-mini pass@1 from 72.7% to 81.9% on Tau2-Bench Telecom.

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 · 2026-10-01

The authors' main claim, as the classifier picked it out 0.54

“Experiments demonstrate improvements on AppWorld and Tau2-Bench.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.06 · states limitation 0.12 · human participants 0.03 · controlled comparison 0.56 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.85 · mentions code or data availability 0.04

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
Language agents are expected to solve increasingly complex tasks, creating a growing need for continual improvement. background 1.00 0.05 0.02 asserted 0.70
One promising approach is to evolve the agent harness, the software that governs tool use, memory management, and task execution, while keeping the underlying language model fixed. background 0.85 0.07 0.06 hedged possibility 0.86
Recent methods automate this process by using a meta coding agent to modify the harness based on execution feedback. background 0.99 0.06 0.03 asserted 1.00
However, relying on that agent's existing knowledge and observed failures can restrict exploration and make adaptation reactive. gap 0.85 0.10 0.03 asserted 0.55
Inspired by how human experts learn from the research literature for new solutions, we introduce ScholarEvolve, a framework that automatically draws on state-of-the-art research to guide harness evolution. contribution 1.00 0.86 0.17 asserted 1.00
ScholarEvolve organizes the harness evolution directions into functional modules and uses topic modeling to identify distinct improvement strategies for each module. method 0.97 0.87 0.06 asserted 1.00
It implements these strategies and evaluates their combinations to improve task performance. method 0.98 0.88 0.08 asserted 0.99
Moreover, the framework is designed to incorporate new publications over time, allowing research advances to drive proactive lifelong evolution. contribution 0.59 0.58 0.10 asserted 1.00
Experiments demonstrate improvements on AppWorld and Tau2-Bench. result 0.99 0.93 0.24 asserted 1.00
ScholarEvolve raises Qwen3.5-27B task goal completion from 49.6% to 63.6% on AppWorld Challenge, and raises GPT-5.4-mini pass@1 from 72.7% to 81.9% on Tau2-Bench Telecom. result 1.00 0.94 0.38 asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.99

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, 2026-10-01: 1 request, 7,502 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 (9)
Figure 1 of the paper
Figure 1. Overview of SCHOLAREVOLVE. Left: workflow focused exploration. Middle: research guided module mutation, recombination, and lifelong evolution. Right: Qwen3.5-27B task (TGC) and scenario (SGC) goal completion on AppWorld Challenge. Recent evaluations document search that plateaus around repeated local edits (Huang et al., 2026). For example, adding verification rounds or multiple agent proposals can increase computation while leaving other mechanisms for improving the harness unexplored (Cemri et al., 2026; Kim et al., 2025)
Figure 2 of the paper
Figure 2. SCHOLAREVOLVE: (a) research guided module mutation, (b) selection and crossover across generations, and (c) evolved modules operating around a fixed backbone. The fixed backbone fθ is the language model whose parameters remain unchanged during evolution. For a task x ∼ Ptar, harness H induces an interaction trajectory τ = (o0, a0, . . . , oT ) containing observations, model decisions, tool calls, and environment responses. We write τ ∼ pθ(· | H, x) and score it by reward R(x, τ ). Harness evolution seeks H⋆ ∈ arg max H∈H J(H), J(H) = Ex∼Ptar Eτ∼pθ(·|H,x) [R(x, τ )]. (1) Starting fro
Figure 3 of the paper
Figure 3. Evolution over time and across held-out tasks. Left: best harness found in each gener￾ation on AppWorld Normal, with bands showing one standard deviation. Right: task-level success￾count transitions pooled over Normal and Challenge (n = 585 per backbone)
Figure 4 of the paper
Figure 4. Residual Qwen3.5-27B failure taxon￾omy with five families and 16 subtypes. Residual failures after evolution. Auditing all 1,336 failed Qwen episodes yields five fail￾ure families and 16 subtypes
Figure 5 of the paper
Figure 5. Reference API distributions across AppWorld splits. Bars show the percentage of tasks at each integer API count, with shared axes across the complete splits. The shaded regions mark at most nine APIs and more than 12 APIs, using the training third quartile and maximum. Dashed lines separate these ranges at integer-bin boundaries. A.2 LIFELONG EVOLUTION PROTOCOL AND RESULTS
Figure 6 of the paper
Figure 6. Reliability transitions by AppWorld test split. Columns and rows give initial and evolved success counts out of three. Cells report task counts. Teal denotes more successes, slate the same count, and rose fewer successes. Darker shades indicate larger fractions of each split on a shared square-root scale. The gains in
Figure 7 of the paper
Figure 7. A hierarchy of observable agent failures. Five families organize evidence-linked devi￾ations found in Qwen3.5-27B AppWorld trajectories. Each subtype identifies a distinct behavior or execution outcome. Labels can coexist within an episode. Recovery status is recorded separately. An execution error is identified by the benchmark’s explicit Execution failed. Traceback: sentinel. Recovery statistics report final success among episodes that encounter this sentinel. The initial and evolved harnesses can encounter errors on different tasks, so these conditional rates describe their…
Figure 8 of the paper
Figure 8. Completion and answer diagnostics on the full Qwen cohorts. Points show fractions of all episodes, and counts give initial → evolved totals. The four endpoint categories follow the diagnostic order from top to bottom. Behavior labels in
Figure 9 of the paper
Figure 9. Implemented mechanisms span the agent’s decision process. Schematic organization of alternatives from the Qwen AppWorld candidate pool. Each module lists two research-derived mechanisms. Arrows show information and action flow through the harness. The alternatives are candidates for selection and recombination. ness creates a new object when the request requires updating an existing one, whereas the evolved harness locates and modifies the original object. In another, an operation unintentionally includes pre-existing state, whereas the evolved harness clears that state before…

Figures are the authors', reproduced from the paper.

References (49)
  1. [1] Lakshya A Agrawal, Shangyin Tan, Dilara Soylu, Noah Ziems, Rishi Khare, Krista Opsahl-Ong, Arnav Singhvi, Herumb Shandilya, Michael J Ryan, Meng Jiang, et al. Gepa: Reflective prompt evolution can outperform reinforcement learning. InInternational Conference on Learning Rep- r... on Pith
  2. [3] Muhammad Asif Ali, Zhengping Li, Shu Yang, Keyuan Cheng, Yang Cao, Tianhao Huang, Guimin Hu, Weimin Lyu, Lijie Hu, Lu Yu, et al. Prompt-saw: Leveraging relation-aware graphs for textual prompt compression.arXiv preprint arXiv:2404.00489, on PithPith's reviewarXiv
  3. [5] Le Chen, Zishen Wan, Baixi Sun, Xiaolong Ma, Chih-Hsuan Yang, Feng Yan, Sheng Di, Franck Cappello, and Rajeev Thakur. Measure before you manage: Evaluating agent working memory in coding agents.arXiv preprint arXiv:2608.31057, on PithPith's reviewarXiv
  4. [6] Dai Do, Quan Tran, Svetha Venkatesh, and Hung Le. Large language model prompting with episodic memory. InECAI 2024: 27th European Conference on Artificial Intelligence, 19–24 October 2024, Santiago de Compostela, Spain–Including 13th Conference on Prestigious Applications of I... on Pith
  5. [7] Ryan Ehrlich, Bradley Brown, Jordan Juravsky, Ronald Clark, Christopher R ´e, and Azalia Mirho- seini. Codemonkeys: Scaling test-time compute for software engineering.arXiv preprint arXiv:2501.14723, on PithPith's reviewarXiv
  6. [9] Juraj Gottweis, Wei-Hung Weng, Alexander Daryin, Tao Tu, Anil Palepu, Petar Sirkovic, Artiom Myaskovsky, Felix Weissenberger, Keran Rong, Ryutaro Tanno, et al. Towards an ai co-scientist. arXiv preprint arXiv:2502.18864, 2, on PithPith's reviewarXiv
  7. [10] Edouardo Honig, Andrew Lizarraga, Zijun Frank Zhang, and Ying Nian Wu. Better prompt com- pression without multi-layer perceptrons.arXiv preprint arXiv:2501.06730, on PithPith's reviewarXiv
  8. [11] 10 Shengran Hu, Cong Lu, and Jeff Clune. Automated design of agentic systems. InInternational Conference on Learning Representations, volume 2025, pp. 21344–21377, on Pith
  9. [12] Lisheng Huang, Chen Yang, Hao Zhou, Huatong Song, Zongchao Chen, Ran Le, Yang Song, Wayne Xin Zhao, and Tao Zhang. Evo-bench: Can language models improve agent harness? arXiv preprint arXiv:2608.09096, on PithPith's reviewarXiv
  10. [13] Seth Karten, Joel Zhang, Tersoo Upaa Jr, Ruirong Feng, Wenzhe Li, Chengshuai Shi, Chi Jin, and Kiran V odrahalli. Continual harness: Online adaptation for self-improving foundation agents. arXiv preprint arXiv:2605.09998, on PithPith's reviewarXiv
  11. [14] Yubin Kim, Ken Gu, Chanwoo Park, Chunjong Park, Samuel Schmidgall, A Ali Heydari, Yao Yan, Zhihan Zhang, Yuchen Zhuang, Yun Liu, et al. Towards a science of scaling agent systems.arXiv preprint arXiv:2512.08296, on PithPith's reviewarXiv
  12. [15] Yoonho Lee, Roshen Nair, Qizheng Zhang, Kangwook Lee, Omar Khattab, and Chelsea Finn. Meta- harness: End-to-end optimization of model harnesses.arXiv preprint arXiv:2603.28052, on PithPith's reviewarXiv
  13. [16] Pietro Lesci, Yoshinari Fujinuma, Momchil Hardalov, Chao Shang, Yassine Benajiba, and Lluis Marquez. Diable: Efficient dialogue state tracking as operations on tables. InFindings of the Association for Computational Linguistics: ACL 2023, pp. 9697–9719, on Pith
  14. [18] Jiahang Lin, Shichun Liu, Chengjun Pan, Lizhi Lin, Shihan Dou, Zhiheng Xi, Xuanjing Huang, Hang Yan, Zhenhua Han, Tao Gui, et al. Agentic harness engineering: Observability-driven automatic evolution of coding-agent harnesses.arXiv preprint arXiv:2604.25850, on PithPith's reviewarXiv
  15. [19] Haozhe Liu, Tian Ye, Sensen Gao, Qihang Cao, Yitong Li, Mingchen Zhuge, Duomin Wang, Ruihua Zhang, Ping Luo, Jiawang Bian, et al. Sol-pi: Recursively scaling auto-research loops for efficient agent harness.arXiv preprint arXiv:2609.20519, 2026a. Xiao Liu, Da Yin, Zirui Wu, and... on PithPith's reviewarXiv
  16. [20] Kexin Ma, Bojun Li, Yuhua Tang, Liting Sun, and Ruochun Jin. Cast: Character-and-scene episodic memory for agents.arXiv preprint arXiv:2602.06051, on PitharXiv
  17. [21] Mike Merrill, Alexander Shaw, Nicholas Carlini, Boxuan Li, Harsh Raj, Ivan Bercovich, Lin Shi, Jeong Shin, Thomas Walshe, E Kelly Buchanan, et al. Terminal-bench: Benchmarking agents on hard, realistic tasks in command line interfaces. InInternational Conference on Learning Re... on Pith
  18. [22] Xuying Ning, Katherine Tieu, Dongqi Fu, Tianxin Wei, Zihao Li, Yuanchen Bei, Jiaru Zou, Mengt- ing Ai, Zhining Liu, Ting-Wei Li, et al. Code as agent harness.arXiv preprint arXiv:2605.18747, on PithPith's reviewarXiv
  19. [23] Siru Ouyang, Jun Yan, I Hsu, Yanfei Chen, Ke Jiang, Zifeng Wang, Rujun Han, Long Le, Samira Daruki, Xiangru Tang, et al. Reasoningbank: Scaling agent self-evolving with reasoning memory. InInternational Conference on Learning Representations, volume 2026, pp. 94327–94354, on Pith
  20. [24] Aditya Ajay Phalod. When should a language model trust itself? same-model self-verification as a conditional confidence signal.arXiv preprint arXiv:2605.02915, on PithPith's reviewarXiv
show 29 more
  1. [25] Chau Minh Pham, Alexander Miserlis Hoyle, Simeng Sun, Philip Resnik, and Mohit Iyyer. Top- icgpt: A prompt-based topic modeling framework. InProceedings of the 2024 conference of the North American chapter of the Association for Computational Linguistics: human language techno... on Pith
  2. [26] Maxime Robeyns, Martin Szummer, and Laurence Aitchison. A self-improving coding agent.arXiv preprint arXiv:2504.15228, on PithPith's reviewarXiv
  3. [27] Samuel Schmidgall and Michael Moor. Agentrxiv: Towards collaborative autonomous research. arXiv preprint arXiv:2503.18102, on PithPith's reviewarXiv
  4. [28] Yu Shang, Yu Li, Keyu Zhao, Likai Ma, Jiahe Liu, Fengli Xu, and Yong Li. Agentsquare: Automatic llm agent search in modular design space. InInternational Conference on Learning Representa- tions, volume 2025, pp. 3841–3865, on Pith
  5. [29] Prashant Shende and Bradley Camburn. Perceptual self-reflection in agentic physics simulation code generation.arXiv preprint arXiv:2602.12311, on PitharXiv
  6. [30] Kimi Team, Tongtong Bai, Yifan Bai, Yiping Bao, Jianfeng Cai, Xinyuan Cai, Peizhou Cao, Yux- uan Cao, Ziwei Chai, Y Charles, et al. Kimi k3: Open frontier intelligence.arXiv preprint arXiv:2607.24653, on PitharXiv
  7. [31] Xingyao Wang, Boxuan Li, Yufan Song, Frank F Xu, Xiangru Tang, Mingchen Zhuge, Jiayi Pan, Yueqi Song, Bowen Li, Jaskirat Singh, et al. Openhands: An open platform for ai software de- velopers as generalist agents. InInternational Conference on Learning Representations, volume ... on Pith
  8. [32] 12 Tianxin Wei, Zhan Shi, Minhua Lin, Bing He, Zewen Liu, Yisi Sang, Yuanchen Bei, Xuying Ning, Jiaru Zou, Ting-Wei Li, et al. Evo-harness: Context-to-harness skill compilation for self-evolving agents.arXiv preprint arXiv:2608.15071, on PithPith's reviewarXiv
  9. [33] Di Wu, Devendra Singh Sachan, Wen-tau Yih, and Mingda Chen. Procedural knowledge at scale improves reasoning.arXiv preprint arXiv:2604.01348, 2026a. George Wu, Nan Jing, Qing Yi, Chuan Hao, Ming Yang, Feng Chang, Yuan Wei, Jian Yang, Ran Tao, and Bryan Dai. Tmas: Scaling test-... on PithPith's reviewarXiv
  10. [34] Yunqing Xu, Xinbei Ma, Jiyang Qiu, and Hai Zhao. Textual-to-visual iterative self-verification for slide generation.arXiv preprint arXiv:2502.15412, on PithPith's reviewarXiv
  11. [35] Yutaro Yamada, Robert Tjarko Lange, Cong Lu, Shengran Hu, Chris Lu, Jakob Foerster, Jeff Clune, and David Ha. The ai scientist-v2: Workshop-level automated scientific discovery via agentic tree search.arXiv preprint arXiv:2504.08066, on PithPith's reviewarXiv
  12. [36] Jingbo Yang, Guanyu Yao, Yang Zhang, Ramana Rao Kompella, Gaowen Liu, and Shiyu Chang. Federatedskill: Federated learning for agentic skill evolution.arXiv preprint arXiv:2606.03143, 2026a. Kaiyue Yang, Yuyan Bu, Jingwei Yi, Yuchi Wang, Biyu Zhou, Juntao Dai, Songlin Hu, and Y... on PithPith's reviewarXiv
  13. [38] Ziyang Yu and Yuyu Liu. Icpc: In-context prompt compression with faster inference.arXiv preprint arXiv:2501.01625, on PithPith's reviewarXiv
  14. [39] Kun Zeng, Yu Huo, Siyu Zhang, Zi Ye, Yuecheng Zhuo, Haoyue Liu, Yuquan Lu, Junhao Wen, and Xiaoying Tang. Group of skills: Group-structured skill retrieval for agent skill libraries.arXiv preprint arXiv:2605.06978, 2026a. Xingshan Zeng, Weiwen Liu, Xu Huang, Zezhong Wang, Ling... on PithPith's reviewarXiv
  15. [40] 13 Qizheng Zhang, Changran Hu, Shubhangi Upasani, Boyuan Ma, Fenglu Hong, Vamsidhar Kama- nuru, Jay Rainton, Chen Wu, Mengmeng Ji, Hanchen Li, et al. Agentic context engineering: Evolving contexts for self-improving language models. InInternational Conference on Learning Repre... on PithPith's reviewarXiv
  16. [41] Huichi Zhou, Siyuan Guo, Anjie Liu, Zhongwei Yu, Ziqin Gong, Bowen Zhao, Zhixun Chen, Men- glong Zhang, Yihang Chen, Jinsong Li, et al. Memento-skills: Let agents design agents.arXiv preprint arXiv:2603.18743, on PitharXiv
  17. [42] 14 APPENDIXCONTENTS A Additional Experimental Details 16 A.1 Implementation Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 A.2 Lifelong Evolution Protocol and Results . . . . . . . . . . . . . . . . . . . . . . . 17 A.3 Training and Test API Distrib... on Pith
  18. [44] +1.2 [−4.1, +6.4] 12/10 Structured API Actions ToolACE-R (Zeng et al., 2026b) +4.7 [−0.6, +9.9] 14/8⋆ Constructed Composite Actions RefTool (Liu et al., 2026b) +1.8 [−4.7, +8.2] 15/10 Context Lossy Context Rewriting Compression without MLPs (Honig et al., 2025)−45.6 [−53.8,−36... on Pith
  19. [45] +9.9 [+4.1, +15.8] 16/5 Adaptive Memory Optimization CoEvo-Mem (Ye et al., 2026)−67.8 [−75.4,−60.2] 0/50 Workflow Evaluator-Guided Refinement Perceptual self-reflection (Shende & Camburn, on Pith
  20. [46] Context Buffer Management Skills Skill-as-Pseudocode (Li et al., 2026)Skill Compilation Skill Drift Is Contract Violation (Fan et al., on Pith
  21. [47] Skill Library Maintenance Procedural KG Extraction (Carriero et al., 2024)Skill Induction Procedural Knowledge at Scale (Wu et al., 2026a) Procedural Memory and Retrieval Memories Oracle Agent Memory (Alake et al., on Pith
  22. [48] Structured Memory Architecture CoEvo-Mem (Ye et al., 2026)Adaptive Retrieval Control CAST (Ma et al., on Pith
  23. [49] Structured Relational Memory MemRL (Zhang et al., 2026d)Adaptive Retrieval and Retention Workflow TMAS (Wu et al., 2026b) Multi-Agent Deliberation Plan-and-Solve (Wang et al., 2023)Plan-and-Execute Textual-to-Visual Self-Verification (Xu et al., 2025)Hierarchical Decomposition... on Pith
  24. [417] Distribution of reference API requirements Share of tasks Figure 5:Reference API distributions across AppWorld splits.Bars show the percentage of tasks at each integer API count, with shared axes across the complete splits. The shaded regions mark at most nine APIs and more th... on Pith
  25. [2019] Bowen Ye, Yongchao Xu, Zhijian Li, Xiang Yin, Junkai Ma, and Wenzhao Li. Coevo-mem: Co- evolving retrieval policy and memory bank for llm agents.arXiv preprint arXiv:2608.01739, on PithPith's reviewarXiv
  26. [2023] Xinze Li, Yuhang Zang, Yixin Cao, and Aixin Sun. Skill-as-pseudocode: Refactoring skill libraries to pseudocode for llm agents.arXiv preprint arXiv:2605.27955, on PithPith's reviewarXiv
  27. [2024] Victor Barres, Honghua Dong, Soham Ray, Xujie Si, and Karthik Narasimhan.τ 2-Bench: Eval- uating conversational agents in a dual-control environment.arXiv preprint arXiv:2506.07982, on PithPith's reviewarXiv
  28. [2025] Linfeng Fan, Yuan Tian, Ziwei Li, and Zhiwu Lu. Skill drift is contract violation: Proactive mainte- nance for llm agent skill libraries.arXiv preprint arXiv:2605.10990, on PithPith's reviewarXiv
  29. [2026] Richmond Alake, Cesare Bernardis, Paul Cayet, Luca Engel, Damien Hilloulin, Sungpack Hong, Allen Hosler, Nickolas Kavantzas, Ingo Kossyk, Son Le, et al. Oracle agent memory as an enterprise memory substrate for long-horizon ai agents.arXiv preprint arXiv:2607.13157, 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