pith. sign in

arxiv: 2605.16309 · v1 · pith:MJQR36JMnew · submitted 2026-05-04 · 💻 cs.AI · cs.LG· cs.MA

ANNEAL: Adapting LLM Agents via Governed Symbolic Patch Learning

Safayat Bin Hakim , Keyan Guo , Wenkai Tan , Alvaro Velasquez , Shouhuai Xu , Houbing Herbert Song This is my paper

Pith reviewed 2026-05-21 00:58 UTC · model grok-4.3

classification 💻 cs.AI cs.LGcs.MA
keywords LLM agentssymbolic patch learningprocess knowledge graphneuro-symbolic systemsfault eliminationgoverned adaptationrecurring failure repair
0
0 comments X

The pith

ANNEAL enables LLM agents to permanently eliminate recurring faults through governed symbolic edits to a process knowledge graph.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that LLM-based agents recover from isolated errors but keep failing on the same issues because their underlying symbolic process knowledge stays unrepaired. ANNEAL introduces Failure-Driven Knowledge Acquisition to locate the faulty operator, generate a constrained patch, and validate it with scoring, guardrails, and canary testing before committing the change with full provenance and rollback. A reader would care if this structural repair approach succeeds where prompt updates and weight changes do not, because it targets the root symbolic structures that encode task execution across repeated runs.

Core claim

ANNEAL converts recurring failures into governed symbolic edits of a process knowledge graph without changing foundation model weights. Its Failure-Driven Knowledge Acquisition mechanism localizes the responsible operator, synthesizes a typed patch through constrained LLM generation, and validates the proposal via multi-dimensional scoring, symbolic guardrails, and canary testing before commit. Across four domains and 27 multi-seed runs, this produces persistent structural repairs that reduce holdout failure rates on recurring faults to 0 percent, while baselines retain 72-100 percent failure rates, and removing the mechanism eliminates all repairs and drops success by up to 26.7 points.

What carries the argument

Failure-Driven Knowledge Acquisition (FDKA), which identifies faulty operators in the process knowledge graph, produces constrained symbolic patches, and commits only those that pass validation with provenance and rollback.

If this is right

  • Agents achieve zero holdout failure on recurring faults through structural changes rather than episodic recovery.
  • All accepted edits carry full provenance and support deterministic rollback.
  • Removing the patch mechanism eliminates structural repairs and reduces success rates by up to 26.7 points.
  • The approach operates across four domains without modifying model weights.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Governed symbolic repair could extend to other knowledge structures such as constraint sets or operator hierarchies.
  • The validation pipeline may need domain-specific strengthening when applied to safety-critical tasks.
  • Combining patch-based edits with occasional weight updates could produce agents that both adapt quickly and retain long-term fixes.

Load-bearing premise

Multi-dimensional scoring, symbolic guardrails, and canary testing together guarantee that every synthesized patch is correct and safe before it is committed.

What would settle it

A holdout test instance containing the original recurring fault that still fails after a patch has been committed to the knowledge graph.

Figures

Figures reproduced from arXiv: 2605.16309 by Alvaro Velasquez, Houbing Herbert Song, Keyan Guo, Safayat Bin Hakim, Shouhuai Xu, Wenkai Tan.

Figure 1
Figure 1. Figure 1: ANNEAL adaptation cycle. User requests hotel booking; execution fails due to policy violation. FDKA localizes the responsible operator, synthesizes a precondition patch, validates through scoring and guardrails, and commits. Replan with updated operator succeeds. Complete cycle: <2 minutes, no model retraining. few of these systems provide the governance guarantees—provenance, guardrails, canary testing, r… view at source ↗
Figure 2
Figure 2. Figure 2: ANNEAL system architecture. Instructions compile into HTN plans via the Process Knowledge Graph (PKG). Metacognitive controller M monitors uncertainty u (token-level entropy) and violation probability pviol (logistic heuristic over precondition gaps) to arbitrate between S1 (fast), S2 (deliberative), and VERIFY (precondition check) pathways. Failures trigger FDKA: localize the responsible operator, synthes… view at source ↗
Figure 3
Figure 3. Figure 3: Adaptation curve (ratchet effect). Cumulative target-class failures versus task index, from real per-task metrics (seed 7; all seeds identical under deterministic failure injec￾tion). (a) Travel stress (12 tasks): ANNEAL patches BookFlight API-drift on first encounter (TTA=0); target failures plateau at 1 while baselines accumulate 9 across prefix and holdout. (b) E-commerce stress (14 tasks): ANNEAL patch… view at source ↗
Figure 4
Figure 4. Figure 4: Concrete PROPOSEEDIT pipeline. Stage 1 serializes failure trace τ into compact JSON. Stage 2 uses few-shot constrained prompting to emit a patch in closed JSON schema. Stage 3 deterministically parses and type-checks output into typed symbolic edit ∆o. The LLM acts as a constrained code generator; all acceptance is handled by subsequent scoring and guardrail stages. E.1 Individual Scoring Sub-Equations Pla… view at source ↗
Figure 5
Figure 5. Figure 5: Four-dimensional scoring geometry. Each patch ∆o is evaluated along plausibility (Eq. 11), consistency (Eq. 12), utility (Eq. 13), and risk (Eq. 14). Aggregate score (Eq. 5 in main text) must exceed θ = 0.18 for acceptance. F Governance Details F.1 Provenance, Rollback, Canary, HITL Gate, Trust Score, and Commit Equations Provenance tuple. Every committed change stores: prov(∆o) = ⟨source, inputs, context,… view at source ↗
Figure 6
Figure 6. Figure 6: FDKA patch example. A payment-authorization failure (PAY-401) on the hotel￾booking operator triggers synthesis of a blocked-card precondition. The box above lists the scoring breakdown and governance outcome; the figure shows the end-to-end pipeline from failure trace to committed edit. The full pipeline is illustrated in [PITH_FULL_IMAGE:figures/full_fig_p024_6.png] view at source ↗
Figure 7
Figure 7. Figure 7 [PITH_FULL_IMAGE:figures/full_fig_p024_7.png] view at source ↗
read the original abstract

LLM-based agents can recover from individual execution errors, yet they repeatedly fail on the same fault when the underlying process knowledge--operator schemas, preconditions, and constraints--remains unrepaired. Existing self-evolving approaches address this gap by updating prompts, memory, or model weights, but none directly repair the symbolic structures that encode how tasks are executed, and few provide the governance guarantees required for safe deployment. We introduce ANNEAL, a neuro-symbolic agent that converts recurring failures into governed symbolic edits of a process knowledge graph without modifying foundation model weights. Its core mechanism, Failure-Driven Knowledge Acquisition (FDKA), localizes the responsible operator, synthesizes a typed patch through constrained LLM generation, and validates the proposal via multi-dimensional scoring, symbolic guardrails, and canary testing before commit. Every accepted edit carries full provenance and deterministic rollback capability. Across four domains and 27 multi-seed runs, ANNEAL is the only evaluated system that commits persistent structural repairs--strong baselines such as ReAct and Reflexion achieve high episodic recovery yet retain 72-100% holdout failure rates on recurring faults, whereas ANNEAL reduces these to 0% in the tested recurring-failure settings. Ablation confirms that removing FDKA eliminates all structural repairs and drops success rate by up to 26.7 percentage points. These results suggest that governed symbolic repair offers a complementary paradigm to weight-level and prompt-level adaptation for persistent fault elimination.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The manuscript introduces ANNEAL, a neuro-symbolic LLM agent framework that performs governed symbolic patch learning on a process knowledge graph. Its core component, Failure-Driven Knowledge Acquisition (FDKA), localizes faulty operators from recurring execution errors, synthesizes typed patches via constrained LLM generation, and validates proposals using multi-dimensional scoring, symbolic guardrails, and canary testing prior to committing edits with full provenance and deterministic rollback. The central empirical claim is that, across four domains and 27 multi-seed runs, ANNEAL is the only evaluated system to commit persistent structural repairs, reducing holdout failure rates on recurring faults to 0% while baselines (ReAct, Reflexion) retain 72-100% failure rates; an ablation removing FDKA eliminates all structural repairs and drops success by up to 26.7 percentage points.

Significance. If the validation pipeline reliably ensures patch correctness and generalization, the work provides a concrete, governance-aware alternative to prompt or weight adaptation for eliminating persistent faults in LLM agents. The quantitative separation from strong baselines on recurring faults, combined with the ablation and multi-seed design, would constitute a notable contribution to neuro-symbolic agent adaptation. The built-in rollback and provenance features further strengthen the case for safe deployment.

major comments (1)
  1. [Abstract] Abstract: The headline result of 0% holdout failure after structural repair rests on the assumption that multi-dimensional scoring, symbolic guardrails, and canary testing together guarantee patch correctness and safety. The abstract provides no specification of the scoring dimensions, the exact preconditions and constraints encoded in the guardrails for each of the four domains, or the construction and coverage of the canary tests (beyond recurring fault instances). Without these details it is impossible to determine whether an incorrect patch could still pass validation and be persisted, undermining the zero-failure claim.
minor comments (1)
  1. [Evaluation] The manuscript would benefit from an explicit enumeration of rejected patch proposals and any post-commit regressions observed on non-recurring tasks, even if only in supplementary material.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their positive assessment of ANNEAL's significance and for the recommendation of major revision. We address the single major comment below with a targeted revision to the abstract.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The headline result of 0% holdout failure after structural repair rests on the assumption that multi-dimensional scoring, symbolic guardrails, and canary testing together guarantee patch correctness and safety. The abstract provides no specification of the scoring dimensions, the exact preconditions and constraints encoded in the guardrails for each of the four domains, or the construction and coverage of the canary tests (beyond recurring fault instances). Without these details it is impossible to determine whether an incorrect patch could still pass validation and be persisted, undermining the zero-failure claim.

    Authors: We agree that the abstract, being a high-level summary, does not enumerate the scoring dimensions, domain-specific guardrail constraints, or canary test construction details. These elements are fully specified in the body of the manuscript (multi-dimensional scoring in Section 3.3, symbolic guardrails encoding preconditions and type constraints per domain in Section 3.2, and canary test construction using recurring faults plus held-out generalization cases in Section 4.2). To directly address the concern and make the zero-failure claim more transparent in the abstract itself, we have revised the abstract to briefly note the scoring dimensions (correctness, safety, and efficiency), indicate that guardrails enforce domain-specific preconditions and constraints, and clarify that canary tests cover both recurring fault instances and additional edge cases for generalization. We believe this change strengthens the presentation while preserving the abstract's conciseness. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical results independent of generation process

full rationale

The paper describes a neuro-symbolic system (ANNEAL/FDKA) that performs governed symbolic edits on a process knowledge graph, followed by empirical evaluation on four domains using 27 multi-seed runs and holdout recurring faults. Success is measured by post-commit failure rates on those holdouts, which are defined externally to the patch synthesis, multi-dimensional scoring, guardrails, and canary testing steps. No equations or claims reduce a reported result to a fitted parameter or self-citation by construction; the central result (0% vs. 72-100% holdout failure) is an observed outcome against independent test instances rather than a renaming or re-derivation of the input mechanisms. The evaluation design is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 1 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities beyond the high-level description of FDKA and the knowledge graph; ledger left minimal.

invented entities (1)
  • Failure-Driven Knowledge Acquisition (FDKA) no independent evidence
    purpose: Localize faulty operator and synthesize validated symbolic patch from recurring failures
    Core mechanism introduced in the abstract as the method for converting failures into governed edits.

pith-pipeline@v0.9.0 · 5817 in / 1265 out tokens · 57982 ms · 2026-05-21T00:58:34.874980+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

30 extracted references · 30 canonical work pages

  1. [1]

    Neuro-symbolic AI+ agent systems: a first reflection on trends, opportunities and challenges , year =

    Belle, Vaishak and Fisher, Michael and Russo, Alessandra and Komendantskaya, Ekaterina and Nottle, Alistair , booktitle =. Neuro-symbolic AI+ agent systems: a first reflection on trends, opportunities and challenges , year =

    work page

  2. [2]

    A neurosymbolic cognitive architecture framework for handling novelties in open worlds , volume =

    Goel, Shivam and Lymperopoulos, Panagiotis and Thielstrom, Ravenna and Krause, Evan and Feeney, Patrick and Lorang, Pierrick and Schneider, Sarah and Wei, Yichen and Kildebeck, Eric and Goss, Stephen and others , journal =. A neurosymbolic cognitive architecture framework for handling novelties in open worlds , volume =

    work page

  3. [3]

    Reflexion: Language Agents with Verbal Reinforcement Learning , year =

    Shinn, Noah and Cassano, Federico and Gopinath, Ashwin and Narasimhan, Karthik and Yao, Shunyu , booktitle =. Reflexion: Language Agents with Verbal Reinforcement Learning , year =

    work page

  4. [4]

    2025 , url =

    Wu, Rong and Wang, Xiaoman and Mei, Jianbiao and Cai, Pinlong and Fu, Daocheng and Yang, Cheng and Wen, Licheng and Yang, Xuemeng and Shen, Yufan and Wang, Yuxin and Shi, Botian , title =. 2025 , url =

    work page 2025

  5. [5]

    MemoryBank: Enhancing Large Language Models with Long-Term Memory , year =

    Zhong, Wanjun and Guo, Lianghong and Gao, Qiqi and Ye, He and Wang, Yanlin , booktitle =. MemoryBank: Enhancing Large Language Models with Long-Term Memory , year =

    work page

  6. [6]

    2026 , url =

    Zhang, Shengtao and Wang, Jiaqian and Zhou, Ruiwen and Liao, Junwei and Feng, Yuchen and Li, Zhuo and Zheng, Yujie and Zhang, Weinan and Wen, Ying and Li, Zhiyu and Xiong, Feiyu and Qi, Yutao and Tang, Bo and Wen, Muning , title =. 2026 , url =

    work page 2026

  7. [7]

    2025 , url =

    He, Yufei and Liu, Juncheng and Liu, Yue and Li, Yibo and Cao, Tri and Hu, Zhiyuan and Xu, Xinxing and Hooi, Bryan , title =. 2025 , url =

    work page 2025

  8. [8]

    Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models , year =

    Zhang, Qizheng and Hu, Changran and Upasani, Shubhangi and Ma, Boyuan and Hong, Fenglu and Kamanuru, Vamsidhar and Rainton, Jay and Wu, Chen and Ji, Mengmeng and Li, Hanchen and Thakker, Urmish and Zou, James and Olukotun, Kunle , booktitle =. Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models , year =

    work page

  9. [9]

    Towards cognitive ai systems: a survey and prospective on neuro-symbolic ai , year =

    Wan, Zishen and Liu, Che-Kai and Yang, Hanchen and Li, Chaojian and You, Haoran and Fu, Yonggan and Wan, Cheng and Krishna, Tushar and Lin, Yingyan and Raychowdhury, Arijit , journal =. Towards cognitive ai systems: a survey and prospective on neuro-symbolic ai , year =

    work page

  10. [10]

    2025 , url =

    Shao, Shuai and Ren, Qihan and Qian, Chen and Wei, Boyi and Guo, Dadi and Yang, Jingyi and Song, Xinhao and Zhang, Linfeng and Zhang, Weinan and Liu, Dongrui and Shao, Jing , title =. 2025 , url =

    work page 2025

  11. [11]

    Fast and accurate task planning using neuro-symbolic language models and multi-level goal decomposition , year =

    Kwon, Minseo and Kim, Yaesol and Kim, Young J , booktitle =. Fast and accurate task planning using neuro-symbolic language models and multi-level goal decomposition , year =

    work page

  12. [12]

    Fast, slow, and metacognitive thinking in AI , volume =

    Bergamaschi Ganapini, M and Campbell, M and Fabiano, F and Horesh, L and Lenchner, J and Loreggia, A and Mattei, N and Rossi, F and Srivastava, B and Venable, KB , journal =. Fast, slow, and metacognitive thinking in AI , volume =

    work page

  13. [13]

    Language Models Are Capable of Metacognitive Monitoring and Control of Their Internal Activations , year =

    Ji-An, Li and Xiong, Hua-Dong and Wilson, Robert C and Mattar, Marcelo G and Benna, Marcus K , booktitle =. Language Models Are Capable of Metacognitive Monitoring and Control of Their Internal Activations , year =

    work page

  14. [14]

    Causal neurosymbolic ai: A synergy between causality and neurosymbolic methods , volume =

    Jaimini, Utkarshani and Henson, Cory and Sheth, Amit , journal =. Causal neurosymbolic ai: A synergy between causality and neurosymbolic methods , volume =

    work page

  15. [15]

    Cognitive Neurosymbolic Artificial Intelligence for Complex Decision-Making: Integrating Foundation Models, Cognitive Architectures, and Knowledge , volume =

    Zi, Yuxin and Roy, Kaushik and Sheth, Amit , journal =. Cognitive Neurosymbolic Artificial Intelligence for Complex Decision-Making: Integrating Foundation Models, Cognitive Architectures, and Knowledge , volume =

    work page

  16. [16]

    2025 , url =

    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 Wang, Hongru and Xiao, Han and Zhou, Yuhang and Zhang, Shaokun and Zhang, Jiayi and Xiang, Jinyu and Fang, Yixiong and Zhao, Qiwen and Liu, Dongrui and Ren, Qihan and Qian, Cheng and Wang, Zhenh...

    work page 2025

  17. [17]

    2026 , url =

    Lin, Minhua and Lu, Hanqing and Shi, Zhan and He, Bing and Mao, Rui and Zhang, Zhiwei and Wu, Zongyu and Tang, Xianfeng and Liu, Hui and Dai, Zhenwei and Zhang, Xiang and Wang, Suhang and Dumoulin, Benoit and Pei, Jian , title =. 2026 , url =

    work page 2026

  18. [18]

    SE-Agent: Self-Evolution Trajectory Optimization in Multi-Step Reasoning with

    Lin, Jiaye and Guo, Yifu and Han, Yuzhen and Hu, Sen and Ni, Ziyi and Wang, Licheng and Chen, Mingguang and Liu, Hongzhang and Chen, Ronghao and He, Yangfan and Jiang, Daxin and Jiao, Binxing and Hu, Chen and Wang, Huacan , booktitle =. SE-Agent: Self-Evolution Trajectory Optimization in Multi-Step Reasoning with. 2025 , url =

    work page 2025

  19. [19]

    2025 , url =

    Sun, Zeyi and Liu, Ziyu and Zang, Yuhang and Cao, Yuhang and Dong, Xiaoyi and Wu, Tong and Lin, Dahua and Wang, Jiaqi , title =. 2025 , url =

    work page 2025

  20. [20]

    Should We Really Edit Language Models? On the Evaluation of Edited Language Models , volume =

    Li, Qi and Liu, Xiang and Tang, Zhenheng and Dong, Peijie and Li, Zeyu and Pan, Xinglin and Chu, Xiaowen , journal =. Should We Really Edit Language Models? On the Evaluation of Edited Language Models , volume =

    work page

  21. [21]

    Wise: Rethinking the knowledge memory for lifelong model editing of large language models , volume =

    Wang, Peng and Li, Zexi and Zhang, Ningyu and Xu, Ziwen and Yao, Yunzhi and Jiang, Yong and Xie, Pengjun and Huang, Fei and Chen, Huajun , journal =. Wise: Rethinking the knowledge memory for lifelong model editing of large language models , volume =

    work page

  22. [22]

    OneEdit: A Neural-Symbolic Collaboratively Knowledge Editing System , url =

    Zhang, Ningyu and Xi, Zekun and Luo, Yujie and Wang, Peng and Tian, Bozhong and Yao, Yunzhi and Zhang, Jintian and Deng, Shumin and Sun, Mengshu and Liang, Lei and others , booktitle =. OneEdit: A Neural-Symbolic Collaboratively Knowledge Editing System , url =

    work page

  23. [23]

    Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik and Cao, Yuan , booktitle =

    work page

  24. [24]

    The Future of Continual Learning in the Era of Foundation Models: Three Key Directions , year =

    Bell, Jack and Quarantiello, Luigi and Coleman, Eric Nuertey and Li, Lanpei and Li, Malio and Madeddu, Mauro and Piccoli, Elia and Lomonaco, Vincenzo , booktitle =. The Future of Continual Learning in the Era of Foundation Models: Three Key Directions , year =

    work page

  25. [25]

    Neuro-Symbolic Task Planning and Replanning using Large Language Models , year =

    Kwon, Minseo and Kim, Young J , booktitle =. Neuro-Symbolic Task Planning and Replanning using Large Language Models , year =

    work page

  26. [26]

    Metacognitive ai: Framework and the case for a neurosymbolic approach , year =

    Wei, Hua and Shakarian, Paulo and Lebiere, Christian and Draper, Bruce and Krishnaswamy, Nikhil and Nirenburg, Sergei , booktitle =. Metacognitive ai: Framework and the case for a neurosymbolic approach , year =

    work page

  27. [27]

    Metagent-P: A Neuro-Symbolic Planning Agent with Metacognition for Open Worlds , year =

    Zhou, Yanfang and Liu, Yuntao and Li, Xiaodong and Zhao, Yongqiang and Wang, Xintong and Tian, Jinlong and Li, Zhenyu and Xu, Xinhai , booktitle =. Metagent-P: A Neuro-Symbolic Planning Agent with Metacognition for Open Worlds , year =

    work page

  28. [28]

    Knowledge Editing for Large Language Models: A Survey , volume =

    Wang, Song and Zhu, Yaochen and Liu, Haochen and Zheng, Zaiyi and Chen, Chen and Li, Jundong , journal =. Knowledge Editing for Large Language Models: A Survey , volume =. 2024 , doi =

    work page 2024

  29. [29]

    CLAUSE: Agentic Neuro-Symbolic Knowledge Graph Reasoning via Dynamic Learnable Context Engineering , year =

    Zhao, Yang and Dai, Chengxiao and Zhuo, Wei and Xiu, Yue and Niyato, Dusit , journal =. CLAUSE: Agentic Neuro-Symbolic Knowledge Graph Reasoning via Dynamic Learnable Context Engineering , year =

    work page

  30. [30]

    Building adaptive knowledge bases for evolving continual learning models , volume =

    Julian, Jack and Koh, Yun Sing and Bifet, Albert , journal =. Building adaptive knowledge bases for evolving continual learning models , volume =

    work page