Pith.

Artificial Intelligence · arXiv:2608.08303 v1 · submitted 8 Aug 2026 · 94 references

Query-Only Backdoor Attacks on Self-Evolving Skills via Trajectory Poisoning

Yuyang Luo, Haoran Wang, Kai Shu

The paper says

“Agentic skills improve large language model (LLM) agents by encoding reusable procedures for complex tasks. However, manually authored skills often adapt poorly to long-horizon tasks and changing environments.”

more of the abstract

To address the limitation, self-evolving skill systems have been developed to automatically construct and update skills from execution trajectories, shifting skill acquisition from external marketplaces to a trusted evolution pipeline. By replacing external skill acquisition with trusted internal construction, self-evolving skill systems reduce exposure to skill injection attacks that rely on direct skill manipulation. However, this skill evolution pipeline may introduce a new attack surface in which an attacker can indirectly steer skill evolution by inducing compromised trajectories through agent interactions. To demonstrate the threat, we propose Trajectory Backdoor Attack (TBA), a query-only attack that steers a trusted skill-evolution pipeline toward producing a backdoored skill. Specifically, we craft attacker-submitted queries to lead the agent to perform the target action and explicitly state the corresponding activation condition in the trajectory. We repeat the same condition-action pattern across diverse triggered tasks, while leaving clean queries unchanged, encouraging the evolver to consolidate the pattern as a reusable trigger-dependent rule into the evolved skill. Experiments on three benchmarks across two skill-evolution systems using four open- and closed-source backbone models demonstrate that TBA reliably implants conditional backdoors while preserving clean-task utility, matching or even surpassing direct skill injection. The results reveal a critical vulnerability in trajectory-driven skill evolution.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 12 Aug 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 new query-only backdoor channel into self-evolving skill pipelines, with clean experiments and one important untested cross-user claim that a revision should address. 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 94 references say.

The paper says, at Section 3.1 and Section 4.1

The review did not quote this passage.

Pith says: Objection 1 of 4 · major

The threat model explicitly claims applicability to cross-user systems (SkillClaw, FederatedSkill), but all experiments use a 40-trajectory evolution batch that is 100% attacker-constructed (10 base tasks, each contributing two triggered and two clean variants), as described in Appendix 1.2. No experiment mixes attacker trajectories with trajectories from independent benign users, and no evolver variant filters by user identity or requires a rule to recur across distinct users.

more of objection 1

The claim that TBA extends to cross-user aggregation is therefore empirically unsupported. The authors should either add experiments with mixed-user batches and provenance-aware evolver variants, or explicitly scope the empirical claim to per-user systems and present the cross-user discussion as a conjecture.

The paper says, at Section 5 and Appendix 3.3

The review did not quote this passage.

Pith says: Objection 2 of 4 · major

The conclusion states that TBA 'evades skill scanners,' but no skill-scanner results appear anywhere in the paper. Appendix 3.3 describes the protocol and says 'We report the number flagged as malicious out of the nine scanned artifacts,' yet no such numbers are presented. This is a load-bearing claim in the conclusion and must be supported by a results table or removed from the claims.

The paper says, at Section 3.2 and Appendix 2

The review did not quote this passage.

Pith says: Objection 3 of 4 · major

The 'policy externalization' component is described as the executor 'synthesizing' or 'inferring' the rule, but the TBA prompt in Appendix 2 contains the literal rule text: 'POLICY_RULE: IF Workflow-Mode == "archival-audit" THEN append_after(...)' and 'BINDING_CHECK: ...'. The executor is therefore restating a rule already supplied by the attacker rather than inferring it.

more of objection 3

This conflates TBA with the Explicit Rule Exposure baseline, which also states the complete rule, and it weakens the claimed novelty of the externalization mechanism. The paper should describe this component as explicit rule prompting with counterfactual contrast, and re-examine the ablation interpretation accordingly.

The paper says, at Section 4.1 and Tables 1-3

The review did not quote this passage.

Pith says: Objection 4 of 4 · major

All headline results are macro-averages over three seeds with no standard deviations, confidence intervals, or significance tests. The claims 'within 0.6-6.1 percentage points of Direct Skill Injection' and improvements of '4.4-25.0 percentage points' over baselines cannot be assessed for statistical reliability with only three seeds. The authors should report per-seed values or error bars and, where possible, a significance test (e.g., paired comparison over seeds or tasks) for the main comparisons.

4 minor objections

The paper says, at Section 4.3

Pith says: minor objection 1 of 4

The text refers to 'Tab. 3a' when discussing the triggered-trajectory ratio, but the referenced result is the sensitivity plot shown in Fig. 3a.

The paper says, at Section 4.3 and Appendix 3.1

Pith says: minor objection 2 of 4

The ablation naming is inconsistent: the main text lists 'w/o typed policy bridge,' Table 5 uses 'w/o policy externalization,' and Appendix 3.1 uses 'w/o typed policy externalization.' These should be unified.

show 2 more

The paper says, at Abstract and Table 3

Pith says: minor objection 3 of 4

The abstract claims TBA is 'matching or even surpassing direct skill injection,' but Table 3 shows TBA is 6.1 and 0.6 percentage points lower than Direct Skill Injection on SkillOpt and Trace2Skill respectively; 'surpassing' is not supported by the reported numbers.

The paper says, at Section 4.2

Pith says: minor objection 4 of 4

The sentence 'TACC generally matches or exceeds CACC' is too broad: Table 1 contains several cells where CACC exceeds TACC (e.g., SkillOpt with Qwen3-VL on SearchQA, 65.0 vs. 63.3). A more precise characterization would be useful.

What would settle it

Construct a mixed-batch experiment: keep the 40-trajectory evolution budget but vary the number of attacker-influenced trajectories k in {4, 8, 12, 20}, filling the remainder with benign trajectories sampled from a different simulated user. Run TBA on both SkillOpt and Trace2Skill and measure TASR. In addition, run a provenance-filtered variant that only consolidates a conditional rule if it appears in trajectories from at least two distinct user IDs. If TASR falls below roughly 10% in either condition, the cross-user threat claimed in §3.1 is not demonstrated by the current experiments.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 12 Aug 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic partly agrees 1 min 38 s · 15,765 tokens read, 10,698 written
  3. Editor wrote the verdict above 13 s · 16,470 tokens read, 1,078 written
  4. Referee major revision: 4 major objections, 4 minor 48 s · 15,949 tokens read, 5,351 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 read this one. The new thing is real: instead of poisoning memory or reflection, TBA poisons the trajectory-consolidation step itself. The attacker holds no skill access and no evolver access, only submits queries. The mechanism is clever and concrete: counterfactual accepted/rejected records teach the executor when to add the target action, a policy-externalization prompt forces the rule into the trajectory in machine-readable form, and cross-task repetition makes the rule look reusable to the evolver. That combination is what makes the attack work, and the ablations support each piece: removing any one of the three drops TASR by roughly 14 to 19 points.

The experiments are broader than most attack papers. Two evolution systems, three benchmarks, four backbones, plus ablation, persistence, and maturity sweeps. The numbers are consistent: TBA beats the interaction-only baselines by wide margins and lands within a few points of direct skill injection, which is a strong result for a query-only attacker. The persistence result is also notable—TASR does not decay over five benign updates.

The soft spots are real but fixable. First, the central cross-user claim is not tested. Every evolution batch is 100% attacker-constructed: 20 triggered and 20 clean queries, all designed by the attacker. The paper says the threat extends to cross-user systems like SkillClaw and FederatedSkill, but no experiment mixes attacker trajectories with independent benign users' trajectories. If an evolver filters by user identity or requires cross-user corroboration, TBA could collapse. That is a load-bearing gap in the paper as written, not a criticism of the attack in per-user or account-compromise settings. Second, the conclusion claims TBA "evades skill scanners," but the only mention of scanner results is the appendix note that nine skills were scanned and the number flagged "will be reported"—no numbers actually appear. That claim should be removed or supported. Third, the macro-averages over three seeds come without error bars or significance tests; given the effect sizes, I doubt the main comparisons flip, but the reader cannot assess stability.

On the circularity concern: the attack query does contain the target policy text, but that is the attacker's design, not a flaw. The attack is empirical, not a derivation. It would be circular only if the evaluation measured what the prompt already said, and it does not—the held-out evaluation is rule-free.

Bottom line: this deserves a serious referee. A revision that adds cross-user mixing experiments, reports the scanner numbers, and adds variance estimates would make it a solid security contribution. I would take it to reading group and likely cite it once it clears those hurdles.

The strongest objection

The attack's core enabling assumption is that the evolver treats attacker-submitted trajectories as ordinary user evidence and consolidates the repeated conditional rule. The paper claims this threat extends to cross-user systems (§3.1, citing SkillClaw and FederatedSkill), but every experiment (Tables 1–3, 5) uses a 40-trajectory evolution batch that is 100% attacker-constructed: 20 triggered and 20 clean queries all designed by the attacker. No run mixes attacker trajectories with independent benign users' trajectories, and no run tests an evolver that filters by user identity, checks provenance, or requires a rule to recur across distinct users. In a real cross-user aggregator, a single attacker's 40 queries may be a small minority, and the cross-task recurrence that TBA uses to make the rule appear 'reusable' may never be observed as cross-user evidence. If the evolver requires corroboration across users or down-weights single-source trajectories, TBA's TASR would likely collapse; the current experimental design cannot rule this out, so the cross-user portion of the central claim is not empirically supported.

The reader identified the provenance-blind evolver as the weakest assumption. I agree that this is a key enabling condition, but the sharper formulation is that the paper never actually places the attacker in a multi-user environment. Section 4.1 uses a single 40-trajectory batch fully controlled by the attacker, so the 'mixed with ordinary user trajectories B_ord' scenario described in §3.2 is never instantiated. The robustness test in Fig. 3a varies the triggered/clean ratio within an attacker-controlled batch, which is not equivalent to dilution by other users. Consequently, the central claim that TBA threatens cross-user self-evolving systems rests on an untested assumption about how such systems aggregate evidence. This is addressable with additional experiments, so the reader's CONDITIONAL verdict remains appropriate. The missing skill-scanner results, absence of error bars, and the overstatement about surpassing direct skill injection are secondary but should be fixed in revision.

The referee's report

Self-evolving skill systems are a realistic deployment model, and the paper identifies a genuinely new attack surface: indirect skill poisoning through queries alone, without direct skill access. If the empirical results hold, TBA shows that the trust boundary of the evolution pipeline can be crossed, which is an important security finding for agentic AI systems. Strengths include the release of code, detailed prompt templates in the appendix, and evaluation across multiple systems and backbones. However, the manuscript currently overstates its evidence in specific ways: the cross-user generalization is asserted but never tested, the conclusion's claim of evading skill scanners is unsupported by any reported results, and the absence of variance or significance measures makes the quantitative comparisons difficult to assess.

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

Self-evolving skill systems replace external skill marketplaces with a trusted pipeline that distills reusable procedures out of execution trajectories, and that shift is usually taken to shrink the attack surface. The paper argues the opposite: the pipeline itself becomes a new attack surface, because an attacker with ordinary query access can craft task queries that make the trusted executor generate poisoned trajectories, and a trusted evolver will then consolidate the attacker's rule into the skill. The proposed Trajectory Backdoor Attack (TBA) combines counterfactual demonstrations, an instruction that forces the executor to state the inferred rule out loud in its trajectory, and repetition of the same rule across diverse triggered tasks, so the rule looks like reusable skill knowledge. Across two evolution systems and three benchmarks, TBA achieves macro-average triggered attack success of 45.6% and 51.1%, within 0.6 to 6.1 percentage points of direct skill injection, while clean-task accuracy stays near benign-evolution levels. If true, the security rationale for moving skills from external artifacts to trusted internal construction collapses, and trajectory-driven skill evolution needs provenance checks of its own.

What carries the argument. The load-bearing mechanism is the attack query itself, constructed as a conditional-policy context: $\mathcal{C}_{\mathrm{CP}} = \mathcal{D}_{\mathrm{cf}} \parallel \mathcal{I}_{\mathrm{ext}}$. The first piece, $\mathcal{D}_{\mathrm{cf}}$, is a four-record counterfactual validator that shows the executor which responses are accepted and rejected with and without the trigger. The second piece, $\mathcal{I}_{\mathrm{ext}}$, is a policy-externalization instruction that requires the executor to state the inferred rule in a fixed POLICY_RULE and BINDING_CHECK format inside the trajectory before solving the task. Submitted across diverse triggered tasks and interleaved with unchanged clean queries, this package makes the executor's trajectories contain explicit, concrete, cross-task evidence of the attacker-specified conditional rule, so the trusted evolver consolidates it as reusable skill knowledge. The core identity the paper aims to establish is Equation (2): the evolved skill must equal the benign skill on clean queries and equal the benign behavior plus the target action exactly when the trigger is present.

Core claim. On the paper's own terms, the central discovery is that a trusted skill-evolution pipeline can be steered by queries alone: an attacker who can only submit task queries and observe responses can cause the evolver to turn a benign skill into a backdoored skill satisfying $\pi_{s_{\mathrm{poi}}}(q)=\pi_{s_{\mathrm{ben}}}(q)\oplus a^*(x_q)$ when $c^*(q)=1$ and $\pi_{s_{\mathrm{ben}}}(q)$ otherwise. The attack gets the trusted executor to perform the target action and to write the conditional policy explicitly into the trajectory, then repeats the same trigger–action pattern across varied tasks while leaving clean queries untouched, which makes the rule appear reusable during consolidation. The paper reports a macro-average Triggered Attack Success Rate (TASR) of 45.6% under SkillOpt and 51.1% under Trace2Skill, within 0.6 to 6.1 percentage points of an oracle that directly writes the rule into the skill artifact, with clean accuracy comparable to benign evolution. It also reports that the backdoor persists through up to five subsequent benign updates, that stronger backbones often amplify rather than resist the attack, and that skill scanning and safeguard backbones give inconsistent protection.

The premise it rests on. The attack only succeeds if the trusted evolver treats attacker-influenced trajectories as ordinary evidence, consolidating them without checking which user the queries came from or whether an explicitly stated policy rule is grounded in the task's actual requirements.

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

A self-evolving skill system that consolidates trajectories without filtering by originating user inherits the vulnerability; the paper notes isolated per-user systems are reachable through account compromise or session hijacking.

The realistic consequence is silent task-preserving sabotage: in the spreadsheet setting, a phrase such as 'legacy reconciliation' can make the evolved skill redirect payment details while the visible edit still succeeds.

The attacker does not need write access to skill files; ordinary query submission suffices, and the paper shows as few as five triggered trajectories in a forty-trajectory batch push TASR to roughly 39.4%.

Stronger backbones are not inherently safer; the paper reports GPT-5.5 reaching about 70% macro-average TASR on both evolution systems.

Artifact-level defenses are insufficient: the paper reports that skill scanning and safeguard backbones provide limited, and sometimes harmful, protection.

Beyond the paper: any evolver reward that privileges cross-task generality—such as preferring policies that recur across trajectories—is itself the lever the attacker pulls; provenance filtering would blunt TBA but may not remove it if an attacker spreads triggered queries across several accounts.

Beyond the paper: the same recipe could serve as a red-team probe for production pipelines, planting a detectable canary rule to test whether a deployment actually verifies trajectory provenance before evolving skills.

Testable extension: vary the evolver's consolidation prompt, e.g., asking it to reject rules that are not grounded in tool-observable task requirements, and measure whether TASR falls; this would isolate which evolver behavior lets the externalized policy survive.

Beyond the paper: because post-attack benign updates did not erase the backdoor in five rounds, the consolidated rule may be encoded as a stable procedural preference; an open question is how many perturbation or unlearning steps would remove it.

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

Agentic skills improve large language model (LLM) agents by encoding reusable procedures for complex tasks. However, manually authored skills often adapt poorly to long-horizon tasks and changing environments. To address the limitation, self-evolving skill systems have been developed to automatically construct and update skills from execution trajectories, shifting skill acquisition from external marketplaces to a trusted evolution pipeline. By replacing external skill acquisition with trusted internal construction, self-evolving skill systems reduce exposure to skill injection attacks that rely on direct skill manipulation. However, this skill evolution pipeline may introduce a new attack surface in which an attacker can indirectly steer skill evolution by inducing compromised trajectories through agent interactions. To demonstrate the threat, we propose Trajectory Backdoor Attack (TBA), a query-only attack that steers a trusted skill-evolution pipeline toward producing a backdoored skill. Specifically, we craft attacker-submitted queries to lead the agent to perform the target action and explicitly state the corresponding activation condition in the trajectory. We repeat the same condition-action pattern across diverse triggered tasks, while leaving clean queries unchanged, encouraging the evolver to consolidate the pattern as a reusable trigger-dependent rule into the evolved skill. Experiments on three benchmarks across two skill-evolution systems using four open- and closed-source backbone models demonstrate that TBA reliably implants conditional backdoors while preserving clean-task utility, matching or even surpassing direct skill injection. The results reveal a critical vulnerability in trajectory-driven skill evolution.

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 0.97

“The results reveal a critical vulnerability in trajectory-driven skill evolution.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.07 · states limitation 0.18 · human participants 0.04 · controlled comparison 0.75 · randomized assignment 0.02 · states quantitative result 0.40 · compares with prior results 0.92 · mentions code or data availability 0.03

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 skills improve large language model (LLM) agents by encoding reusable procedures for complex tasks. background 1.00 0.05 0.02 asserted 1.00
However, manually authored skills often adapt poorly to long-horizon tasks and changing environments. background 0.91 0.06 0.02 asserted 0.97
To address the limitation, self-evolving skill systems have been developed to automatically construct and update skills from execution trajectories, shifting skill acquisition from external marketplaces to a trusted evolution pipeline. background 1.00 0.06 0.04 asserted 1.00
By replacing external skill acquisition with trusted internal construction, self-evolving skill systems reduce exposure to skill injection attacks that rely on direct skill manipulation. background 0.95 0.08 0.05 asserted 0.96
However, this skill evolution pipeline may introduce a new attack surface in which an attacker can indirectly steer skill evolution by inducing compromised trajectories through agent interactions. gap 0.74 0.13 0.14 hedged possibility 1.00
To demonstrate the threat, we propose Trajectory Backdoor Attack (TBA), a query-only attack that steers a trusted skill-evolution pipeline toward producing a backdoored skill. contribution 0.97 0.90 0.15 asserted 1.00
Specifically, we craft attacker-submitted queries to lead the agent to perform the target action and explicitly state the corresponding activation condition in the trajectory. method 0.99 0.89 0.04 asserted 0.98
We repeat the same condition-action pattern across diverse triggered tasks, while leaving clean queries unchanged, encouraging the evolver to consolidate the pattern as a reusable trigger-dependent rule into the evolved skill. method 1.00 0.91 0.06 asserted 0.69
Experiments on three benchmarks across two skill-evolution systems using four open- and closed-source backbone models demonstrate that TBA reliably implants conditional backdoors while preserving clean-task utility, matching or even surpassing direct skill injection. result 1.00 0.96 0.36 asserted 1.00
The results reveal a critical vulnerability in trajectory-driven skill evolution. interpretation 0.98 0.89 0.71 asserted 1.00

What kind of paper the classifier reports this is: method or tool 0.96 · 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, 2026-09-23: 1 request, 7,594 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 (3)
Figure 1 of the paper
Figure 1. Illustration of backdoor attack in self-evolving skill
Figure 2 of the paper
Figure 2. Overview of the Trajectory Backdoor Attack (TBA). Counterfactual demonstrations and a policy-externalization
Figure 3 of the paper
Figure 3. Sensitivity and robustness of TBA on SkillOpt

Figures are the authors', reproduced from the paper.

References (94)
  1. [1] Jimenez, Carlos E and Yang, John and Wettig, Alexander and Yao, Shunyu and Pei, Kexin and Press, Ofir and Narasimhan, Karthik , title =. International Conference on Learning Representations , year = on Pith
  2. [2] Yang, John and Jimenez, Carlos E and Wettig, Alexander and Lieret, Kilian and Yao, Shunyu and Narasimhan, Karthik and Press, Ofir , title =. 2024 , eprint = on Pith
  3. [3] Deng, Xiang and Gu, Yu and Zheng, Boyuan and Chen, Shijie and Stevens, Sam and Wang, Boshi and Sun, Huan and Su, Yu , title =. Advances in Neural Information Processing Systems , year = on Pith
  4. [4] Lu, Chris and Lu, Cong and Lange, Robert Tjarko and Foerster, Jakob and Clune, Jeff and Ha, David , title =. 2024 , eprint = on Pith
  5. [5] Boiko, Daniil A and MacKnight, Robert and Kline, Ben and Gomes, Gabe , title =. Nature , volume = on Pith
  6. [6] Dunn, Matthew and Sagun, Levent and Higgins, Mike and Guney, V. Ugur and Cirik, Volkan and Cho, Kyunghyun , title =. 2017 , eprint = on Pith
  7. [7] Opsahl-Ong, Krista and Singhvi, Arnav and Collins, Jasmine and Zhou, Ivan and Wang, Cindy and Baheti, Ashutosh and Oertell, Owen and Portes, Jacob and Havens, Sam and Elsen, Erich and Bendersky, Michael and Zaharia, Matei and Chen, Xing , title =. 2026 , eprint = on Pith
  8. [8] He, Linyang and Yu, Qiyao and Dong, Hanze and Liao, Baohao and Xu, Xinxing and Goldblum, Micah and Bian, Jiang and Mesgarani, Nima , title =. 2026 , eprint = on Pith
  9. [9] Ma, Zeyao and Zhang, Bohan and Zhang, Jing and Yu, Jifan and Zhang, Xiaokang and Zhang, Xiaohan and Luo, Sijia and Wang, Xi and Tang, Jie , title =. Advances in Neural Information Processing Systems , volume = on Pith
  10. [10] Bai, Shuai and Cai, Yuxuan and Chen, Ruizhe and others , title =. 2025 , eprint = on Pith
  11. [11] Liu, Aixin and Mei, Aoxue and Lin, Bangcai and Xue, Bing and Wang, Bingxuan and Xu, Bingzheng and Wu, Bochao and Zhang, Bowei and Lin, Chaofan and Dong, Chen and others , title =. 2025 , eprint = on Pith
  12. [12] Agarwal, Sandhini and Ahmad, Lama and Ai, Jason and Altman, Sam and Applebaum, Andy and Arbus, Edwin and Arora, Rahul K and Bai, Yu and Baker, Bowen and Bao, Haiming and others , title =. 2025 , eprint = on Pith
  13. [13] Kimi K2.5: Visual Agentic Intelligence , year =. 2602.02276 , archivePrefix = on PithPith's reviewarXiv
  14. [14] GLM-5: from Vibe Coding to Agentic Engineering , year =. 2602.15763 , archivePrefix = on PithPith's reviewarXiv
  15. [15] OpenAI , title =. 2025 , howpublished = on Pith
  16. [16] OpenAI , title =. 2026 , howpublished = on Pith
  17. [17] Wang, Guanzhi and Xie, Yuqi and Jiang, Yunfan and Mandlekar, Ajay and Xiao, Chaowei and Zhu, Yuke and Fan, Linxi and Anandkumar, Anima , title =. Transactions on Machine Learning Research , year = on Pith
  18. [18] Yutao Yang and Junsong Li and Qianjun Pan and Bihao Zhan and Yuxuan Cai and Lin Du and Jie Zhou and Kai Chen and Qin Chen and Xin Li and Bo Zhang and Liang He , title=. 2026 , eprint= on Pith
  19. [19] Ni, Jingwei and Liu, Yihao and Liu, Xinpeng and Sun, Yutao and Zhou, Mengyu and Cheng, Pengyu and Wang, Dexin and Zhao, Erchao and Jiang, Xiaoxi and Jiang, Guanjun , title =. 2026 , eprint = on Pith
  20. [20] Liu, Xingyan and Luo, Xiyue and Li, Linyu and Huang, Ganghong and Liu, Jianfeng and Qiao, Honglin , title =. 2026 , eprint = on Pith
show 74 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 others , title =. 2026 , eprint = on Pith
  2. [22] Ma, Ziyu and Yang, Shidong and Ji, Yuxiang and Wang, Xucong and Wang, Yong and Hu, Yiming and Huang, Tongwen and Chu, Xiangxiang , title =. 2026 , eprint = on Pith
  3. [23] Yang, Jingbo and Yao, Guanyu and Zhang, Yang and Kompella, Ramana Rao and Liu, Gaowen and Chang, Shiyu , title =. 2026 , eprint = on Pith
  4. [24] Du, Jiawei and Wu, Jinlong and Chen, Yuzheng and Hu, Yucheng and Li, Bing and Zhou, Joey Tianyi , title =. 2025 , eprint = on Pith
  5. [25] Wang, Zora Zhiruo and Mao, Jiayuan and Fried, Daniel and Neubig, Graham , title =. 2024 , eprint = on Pith
  6. [26] OpenSpace: Make Your Agents Smarter, Low-Cost, Self-Evolving , year = on Pith
  7. [27] Huang, Zisu and Xu, Jingwen and Yang, Yifan and Gong, Ziyang and Yang, Qihao and Tian, Muzhao and Wang, Xiaohua and Lv, Changze and Gao, Xuemei and Dai, Qi and Liu, Bei and Qiu, Kai and Yang, Xue and Chen, Dongdong and Zheng, Xiaoqing and Luo, Chong , title =. 2026 , eprint = on Pith
  8. [28] 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 , title =. 2026 , eprint = on Pith
  9. [29] Zheng, Junhao and Shi, Chengming and Cai, Xidi and Li, Qiuke and Zhang, Duzhen and Li, Chenxing and Yu, Dong and Ma, Qianli , title =. IEEE Transactions on Pattern Analysis and Machine Intelligence , publisher = on Pith
  10. [30] Equipping Agents for the Real World with Agent Skills , year = on Pith
  11. [31] Wang, Chenxi and Yu, Zhuoyun and Xie, Xin and Yao, Wuguannan and Fang, Runnan and Qiao, Shuofei and Cao, Kexin and Zheng, Guozhou and Qi, Xiang and Zhang, Peng and others , title =. 2026 , eprint = on Pith
  12. [32] Tu, Songjun and Xu, Chengdong and Zhang, Qichao and Zhang, Yaocheng and Lan, Xiangyuan and Li, Linjing and Zhao, Dongbin , title =. 2026 , eprint = on Pith
  13. [33] Shao, Shuai and Ren, Qianru and Qian, Chen and Wei, Boqun and Guo, Dian and Yang, Jieyu and Song, Xiao and Zhang, Leidong and Liu, Dong and Shao, Jing , title =. 2025 , eprint = on Pith
  14. [34] Jiang, Yanna and Li, Delong and Deng, Haiyu and Ma, Baihe and Wang, Xu and Wang, Qin and Yu, Guangsheng , title =. 2026 , eprint = on Pith
  15. [35] Ma, Yuchen and Huang, Yue and Bao, Han and Zhuang, Haomin and Shukla, Swadheen and Galley, Michel and Zhang, Xiangliang and Feuerriegel, Stefan , title =. 2026 , eprint = on Pith
  16. [36] Yang, Min and Piao, Jinghua and Xia, Xu and Lan, Xiaochong and Chen, Jiaju and Gong, Yongshun and Li, Yong , title =. 2026 , eprint = on Pith
  17. [37] Wang, Hanyu and Lan, Yifan and Cao, Bochuan and Lin, Lu and Chen, Jinghui , title =. 2026 , eprint = on Pith
  18. [38] Li, Yubo , title =. Transactions on Machine Learning Research , year = on Pith
  19. [39] Liu, Yi and Wang, Weizhe and Feng, Ruitao and Zhang, Yao and Xu, Guangquan and Deng, Gelei and Li, Yuekang and Zhang, Leo , title =. 2026 , eprint = on Pith
  20. [40] Li, Zhiyuan and Wu, Jingzheng and Ling, Xiang and Cui, Xing and Luo, Tianyue , title =. 2026 , eprint = on Pith
  21. [41] Gao, Huan-ang and Geng, Jiayi and Hua, Wenyue and Hu, Mengkang and Juan, Xinzhe and Liu, Hongzhang and Liu, Shilong and Qiu, Jiahao and Qi, Xuan and Wu, Yiran and others , title =. 2025 , eprint = on Pith
  22. [42] Fang, Jinyuan and Peng, Yanwen and Zhang, Xi and Wang, Yingxu and Yi, Xinhao and Zhang, Guibin and Xu, Yi and Wu, Bin and Liu, Siwei and Li, Zihao and others , title =. 2025 , eprint = on Pith
  23. [43] Wang, Kaixiang and Lou, Jiong and Zhou, Zhaojiacheng and Li, Jie , title =. 2026 , eprint = on Pith
  24. [44] Wang, Hongtao and Yang, Se and Chen, Yu and Liu, Puzhuo , title =. 2026 , eprint = on Pith
  25. [45] Jia, Xiaojun and Liao, Jie and Qin, Simeng and Gu, Jindong and Ren, Wenqi and Cao, Xiaochun and Liu, Yang and Torr, Philip , title =. 2026 , eprint = on Pith
  26. [46] Qu, Yubin and Liu, Yi and Geng, Tongcheng and Deng, Gelei and Li, Yuekang and Zhang, Leo Yu and Zhang, Ying and Ma, Lei , title =. 2026 , eprint = on Pith
  27. [47] Schmotz, David and Beurer-Kellner, Luca and Abdelnabi, Sahar and Andriushchenko, Maksym , title =. 2026 , eprint = on Pith
  28. [48] Feng, Yunhao and Ding, Yifan and Tan, Yingshui and Zheng, Boren and Guo, Yanming and Li, Xiaolong and Zhai, Kun and Li, Yishan and Huang, Wenke , title =. 2026 , eprint = on Pith
  29. [49] Tie, Guiyao and Shi, Jiawen and Zhou, Pan and Sun, Lichao , title =. 2026 , eprint = on Pith
  30. [50] Ning, Yuting and Zhang, Zhehao and Lal, Yash Kumar and Gou, Boyu and Li, Junyi and Ruan, Weitong and Ye, Chentao and Gupta, Rahul and Yang, Diyi and Su, Yu and others , title =. 2026 , eprint = on Pith
  31. [51] Duan, Zenghao and Tian, Yuxin and Yin, Zhiyi and Pang, Liang and Deng, Jingcheng and Wei, Zihao and Xu, Shicheng and Ge, Yuyao and Cheng, Xueqi , title =. 2026 , eprint = on Pith
  32. [52] Zhou, Zhaojiacheng , title =. 2026 , eprint = on Pith
  33. [53] Liu, Xinyu and Zhao, Yukai and Hu, Xing and Xia, Xin , title =. 2026 , eprint = on Pith
  34. [54] Xu, Renjun and Yan, Yang , title =. 2026 , eprint = on Pith
  35. [55] John Yang and Kilian Lieret and Carlos E Jimenez and Alexander Wettig and Kabir Khandpur and Yanzhe Zhang and Binyuan Hui and Ofir Press and Ludwig Schmidt and Diyi Yang , title =. The Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Benchma... on Pith
  36. [56] Cui, Can and Yang, Zichong and Zhou, Yupeng and Ma, Yunsheng and Lu, Juanwu and Li, Lingxi and Chen, Yaobin and Panchal, Jitesh and Wang, Ziran , title =. 2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC) , pages = on Pith
  37. [57] Jin, Ye and Yang, Ruoxuan and Yi, Zhijie and Shen, Xiaoxi and Peng, Huiling and Liu, Xiaoan and Qin, Jingli and Li, Jiayang and Xie, Jintao and Gao, Peizhong and others , title =. 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages = on Pith
  38. [58] Jiageng Mao and Junjie Ye and Yuxi Qian and Marco Pavone and Yue Wang , title =. First Conference on Language Modeling , year = on Pith
  39. [59] Skill Scanner , year = on Pith
  40. [60] Li, Xiangyi and Chen, Wenbo and Liu, Yimin and Zheng, Shenghan and Chen, Xiaokun and He, Yifeng and Li, Yubo and You, Bingran and Shen, Haotian and Sun, Jiankai and others , title =. 2026 , eprint = on Pith
  41. [61] Guo, Zihan and Chen, Zhiyu and Nie, Xiaohang and Lin, Jianghao and Zhou, Yuanjian and Zhang, Weinan , title =. 2026 , eprint = on Pith
  42. [62] Socket Brings Supply Chain Security to skills.sh , year = on Pith
  43. [63] Ding, Kexin and Zhou, Yang and Jin, Can and Tong, Feng and Zhou, Mu and Metaxas, Dimitris N , title =. 2026 , eprint = on Pith
  44. [64] Lei, Yingtie and Wan, Zhongwei and Zhang, Jiankun and Alam, Samiul and Zhong, Zixuan and Huang, Peizhou and Wang, Xin and Zhang, Jingxuan and Zhou, Donghao and Hsieh, Yunta and others , title =. 2026 , eprint = on Pith
  45. [65] Chen, Zhaorun and Xiang, Zhen and Xiao, Chaowei and Song, Dawn and Li, Bo , title =. Advances in Neural Information Processing Systems , year = on Pith
  46. [66] Dong, Shen and Xu, Shaochen and He, Pengfei and Li, Yige and Tang, Jiliang and Liu, Tianming and Liu, Hui and Xiang, Zhen , title =. Advances in Neural Information Processing Systems , year = on Pith
  47. [67] Srivastava, Saksham Sahai and He, Haoyu , title =. 2025 , eprint = on Pith
  48. [68] Yang, Xianglin and He, Yufei and Ji, Shuo and Hooi, Bryan and Dong, Jin Song , title =. 2026 , eprint = on Pith
  49. [69] Zhong, Wanjun and Guo, Lianghong and Gao, Qiqi and Ye, He and Wang, Yanlin , title =. Proceedings of the AAAI Conference on Artificial Intelligence , year = on Pith
  50. [70] Packer, Charles and Fang, Vivian and Patil, Shishir G and Lin, Kevin and Wooders, Sarah and Gonzalez, Joseph E , title =. 2023 , eprint = on Pith
  51. [71] Ling, George and Zhong, Shanshan and Huang, Richard , title =. 2026 , eprint = on Pith
  52. [72] Shinn, Noah and Cassano, Federico and Gopinath, Ashwin and Narasimhan, Karthik and Yao, Shunyu , title =. Advances in Neural Information Processing Systems , volume = on Pith
  53. [73] Jiayi Zhang and Jinyu Xiang and Zhaoyang Yu and Fengwei Teng and Xiong-Hui Chen and Jiaqi Chen and Mingchen Zhuge and Xin Cheng and Sirui Hong and Jinlin Wang and Bingnan Zheng and Bang Liu and Yuyu Luo and Chenglin Wu , title =. The Thirteenth International Conference on Lear... on Pith
  54. [74] Yuan, Siyu and Song, Kaitao and Chen, Jiangjie and Tan, Xu and Li, Dongsheng and Yang, Deqing , title =. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Pape... on Pith
  55. [75] Lianmin Zheng and Wei-Lin Chiang and Ying Sheng and Siyuan Zhuang and Zhanghao Wu and Yonghao Zhuang and Zi Lin and Zhuohan Li and Dacheng Li and Eric Xing and Hao Zhang and Joseph E. Gonzalez and Ion Stoica , title =. Thirty-seventh Conference on Neural Information Processing... on Pith
  56. [76] Hermes Agent , year = on Pith
  57. [77] Zhen Xiang and Fengqing Jiang and Zidi Xiong and Bhaskar Ramasubramanian and Radha Poovendran and Bo Li , title =. NeurIPS 2023 Workshop on Backdoors in Deep Learning - The Good, the Bad, and the Ugly , year = on Pith
  58. [78] Xu, Jiashu and Ma, Mingyu and Wang, Fei and Xiao, Chaowei and Chen, Muhao , title =. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages = on Pith
  59. [79] Chen, Liuji and Yang, Xiaofang and Lu, Yuanzhuo and Zhang, Jinghao and Sun, Xin and Liu, Qiang and Wu, Shu and Dong, Jing and Wang, Liang , title =. 2025 , eprint = on Pith
  60. [80] Wang, Zhiqiang and Gao, Yichao and Wang, Yanting and Liu, Suyuan and Sun, Haifeng and Cheng, Haoran and Shi, Guanquan and Du, Haohua and Li, Xiangyang , title =. Proceedings of the AAAI Conference on Artificial Intelligence , volume = on Pith
  61. [81] Luo, Junyu and Zhang, Weizhi and Yuan, Ye and Zhao, Yusheng and Yang, Junwei and Gu, Yiyang and Wu, Bohan and Chen, Binqi and Qiao, Ziyue and Long, Qingqing and others , title =. 2025 , eprint = on Pith
  62. [82] Parisi, Aaron and Zhao, Yao and Fiedel, Noah , title =. 2022 , eprint = on Pith
  63. [83] Wujiang Xu and Zujie Liang and Kai Mei and Hang Gao and Juntao Tan and Yongfeng Zhang , title =. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year = on Pith
  64. [84] Olsson, Catherine and Elhage, Nelson and Nanda, Neel and Joseph, Nicholas and DasSarma, Nova and Henighan, Tom and Mann, Ben and Askell, Amanda and Bai, Yuntao and Chen, Anna and Conerly, Tom and Drain, Dawn and Ganguli, Deep and Hatfield-Dodds, Zac and Hernandez, Danny and Jo... on Pith
  65. [85] Liu, Nelson F and Lin, Kevin and Hewitt, John and Paranjape, Ashwin and Bevilacqua, Michele and Petroni, Fabio and Liang, Percy , title =. Transactions of the Association for Computational Linguistics , volume = on Pith
  66. [86] Gao, Tianyu and Fisch, Adam and Chen, Danqi , editor =. Making Pre-trained Language Models Better Few-shot Learners , booktitle =. 2021 , doi = on Pith
  67. [87] Chia, Yew Ken and Chen, Guizhen and Tuan, Luu Anh and Poria, Soujanya and Bing, Lidong , title =. 2023 , eprint = on Pith
  68. [88] Mo, Ying and Liu, Jiahao and Yang, Jian and Wang, Qifan and Zhang, Shun and Wang, Jingang and Li, Zhoujun , editor =. Findings of the Association for Computational Linguistics: EMNLP 2024 , pages =. 2024 , doi = on Pith
  69. [89] Honovich, Or and Shaham, Uri and Bowman, Samuel R. and Levy, Omer , editor =. Instruction Induction: From Few Examples to Natural Language Task Descriptions , booktitle =. 2023 , doi = on Pith
  70. [90] Zhao, Andrew and Huang, Daniel and Xu, Quentin and Lin, Matthieu and Liu, Yong-Jin and Huang, Gao , title =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2024 , doi = on Pith
  71. [91] Share Agents with Other Users , year = on Pith
  72. [92] Analyze Conversational Agents , year = on Pith
  73. [93] Skills System,. 2026 , howpublished = on Pith
  74. [94] Curator,. 2026 , howpublished = on Pith
Cited by 2 Pith papers
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 · 25 sentences without a confident label

Uses the method or tool of 5 papers
Cites as support 2 papers
  • Self-evolving systems therefore shift skill acquisition from external artifacts to trusted internal construction and have demonstrated improved performance and adaptability over static skills [ref] .

    SkillOpt: Executive Strategy for Self-Evolving Agent Skills · Introduction · 0.98

  • Prior work shows that jointly presenting valid and invalid behaviors helps language models distinguish a desired pattern from plausible alternatives [ref] .

    Contrastive Chain-of-Thought Prompting · Counterfactual Trajectory Induction · 0.96

Cites as background 13 papers

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