acceptodds
Withdrawn from ICLR 2027

Knowledge-Centric Self-Improvement

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Self-improving AI systems typically treat the agent as the object that improves, by optimizing prompts, workflows, harnesses, or even the agent’s own code. This agent-centric view can make improvements expensive to maintain and difficult to transfer, because gains become tied to a particular agent design, task distribution, or adaptation run. We study a complementary paradigm: knowledge-centric self-improvement, in which agents remain generic and disposable while the persistent object is a curated knowledge base that agents can leverage for future tasks. We conduct controlled case studies to operationalize this idea via a simple protocol. Agents attempt one task, then contribute evidence-grounded insights to a shared knowledge base via task-level and cross-task forums, followed by knowledge distillation. Because self-improvement is contained in the knowledge rather than the agent, improvement can be more inspectable, transferable, and portable. Across abstract reasoning, coding, and terminal benchmarks, this protocol improves solve rates while reducing dollar cost relative to agent-centric baselines. The resulting distilled knowledge also transfers to held-out tasks and across LLM families, indicating that the improvement is not merely an LLM- or run-specific behavior. These results support a new view of self-improving agentic systems: progress can be driven primarily by the curated persistent knowledge.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM agents, Self improvement

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_343jOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes Knowledge-Centric Self-Improvement (KSI), a framework that improves agent performance by aggregating task experiences through forum-based discussions and distilling them into a shared knowledge base.

Critical strengths

The paper investigates an important research question about agent self-improvement and evaluates its effectiveness across multiple benchmarks, including coding and terminal tasks.

Critical weaknesses
  1. Very low novelty. The work is largely identical to FoT [1], CORAL [2], FederatedSkill [3], SAGE [4], etc. These works are not even discussed.

  2. The experiments do not support the claims. First, removing the cross-task forum only changes performance from 76% to 72%, corresponding to only two tasks on Polyglot. This ablation is also conducted using a single seed. In the main experiments, agents repeatedly attempt the same pool of tasks over ten generations while retaining task-specific knowledge from previous attempts. Consequently, improvements may largely result from accumulating information about previously encountered tasks rather than acquiring transferable knowledge. ARC-AGI-1 and ARC-AGI-2 contain tasks from the same domain, while SWE-bench Pro and Polyglot are both coding benchmarks. They cannot support the cross-domain claim. Terminal-Bench 2 can reflect cross-domain capabilities.

[1] Yao, Dixi, Tahseen Rabbani, Manzil Zaheer, and Tian Li. "Federation over Text: Insight Sharing for Multi-Agent Reasoning." arXiv preprint arXiv:2604.16778 (2026).

[2] Qu, Ao, Han Zheng, Zijian Zhou, Yihao Yan, Yihong Tang, Shao Yong Ong, Fenglu Hong et al. "CORAL: Towards Autonomous Multi-Agent Evolution for Open-Ended Discovery." arXiv preprint arXiv:2604.01658 (2026).

[3] Yang, Jingbo, Guanyu Yao, Yang Zhang, Ramana Rao Kompella, Gaowen Liu, and Shiyu Chang. "FederatedSkill: Federated Learning for Agentic Skill Evolution." arXiv preprint arXiv:2606.03143 (2026).

[4] Pan, Linyue, Yaoming Zhu, Lin Qiu, Xuezhi Cao, and Xunliang Cai. "SAGE: A Quantitative Evaluation of Socialized Evolution in Agent Ecosystems." arXiv preprint arXiv:2606.03544 (2026).

Questions for authors

No particular question.

Reviewer_EmsZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper describes a method where the Agent stores the experience gained during problem-solving into an external knowledge base, which is then used by subsequent Agents. The model and Agent framework remain fixed; only the knowledge is updated. Specific methods include discussions around the same problem, cross-problem exchanges, and the compilation of useful information from these discussions.

Critical strengths

The method is simple and practical, and the accumulated knowledge is easy to check and reuse. Each round doesn't require using the same agent as before; only useful experience needs to be transferred to the new agent. This experience can also be used on different models, and the paper provides results on cross-model transfer.

The authors included a control group of repeated trials, demonstrating that the benefits are not solely due to running the program multiple times. In Appendix N, the complete method achieved 76% success rate in a single run on Polyglot, compared to 62% for the control group that ran the same method for ten rounds but did not accumulate knowledge.

Critical weaknesses

The extent to which it helps with truly new questions hasn't been thoroughly tested. The main experiment involved running the same batch of questions multiple times, counting the cumulative number of questions solved. This shows the system is better at utilizing previous attempts, but it doesn't necessarily mean its performance on new questions has improved. The authors' supplementary experiment with new questions only had 20 questions per benchmark, which is relatively few.

Questions for authors

Compared to CORAL, the new contributions in this paper are not yet clearly explained. CORAL has already enabled multiple agents to accumulate and reuse experience through shared trial logs, notes, and skills. Therefore, similar work already exists on the idea of ​​"storing experience in an external knowledge base for other agents to continue using." This paper adds new agent takeover, intra-task and cross-task discussions, and a unified knowledge organization process, but does not discuss its relationship with CORAL. Although the authors conducted ExpeL, ACE comparisons, and component ablation, it is still necessary to explain: compared to CORAL's direct experience sharing approach, what specific problems does this paper solve, and what new designs can existing experiments prove to be useful?

Reviewer_FAbnOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes keeping task-solving agents generic and disposable and making a curated knowledge base the only persistent, improving object. In each generation, a fresh agent attempts one task with standard SDK tools and seed bundles, then posts schema-constrained, evidence-citing claims to a per-task forum and a cross-task forum, where later rounds must AGREE, DISAGREE or SYNTHESIZE with a cited peer post. An LLM distills the posts into a bundle for each unsolved task and one shared cross-task bundle (Section 3; Appendix F). Over 10 generations and three seeds with Haiku 4.5 and GPT-5.4-mini, on 50-task pools of ARC-AGI-1, ARC-AGI-2, Polyglot and SWE-bench Pro and on all 89 Terminal-Bench 2 tasks, the paper reports higher solve rates at lower dollar cost than rerun DGM and HyperAgents (Table 2a), budget-matched GEPA and OpenEvolve, and in-loop ExpeL and ACE (Table 3), and 43.8% on Terminal-Bench 2 against 37.6% reported for Meta-Harness (Table 2b). A frozen generation-10 knowledge base also improves zero-shot solve rates on 20 held-out Polyglot and ARC-AGI-1 tasks, including across LLM families (Table 5).

Critical strengths
  • An inspectable design with concrete grounding, and a fair in-loop comparison with experience reuse. An agent cannot post until it has called knowledge() and query() on the task (Appendix F.1). Later cross-task posts must cite evidence task ids and take a stance, and bundle items carry applies_when and does_not_apply_when scopes (Appendix F.3). Appendix C traces disagreements into bundles and later solves, for example ARC task b548a754, solved in generation 2 after a round-1 DISAGREE narrowed a falsified hypothesis. Inside the same loop with per-task knowledge added, ExpeL and ACE trail the protocol: 78% and 72% against 86.7% on ARC-AGI-1, and 56% and 50% against 68.0% on Polyglot (Table 3).
  • The held-out transfer study tests the thesis directly. A frozen knowledge base is applied to disjoint held-out tasks; for ARC-AGI-1 the held-out tasks come from the evaluation split and the curation pool from the training split. Recipients never update the knowledge, all donor–recipient pairs across both LLM families are run, and a rerun no-knowledge reference accounts for selecting tasks that both models failed. Mean solve rates improve in all eight combinations, most on ARC-AGI-1, where the GPT-5.4-mini recipient goes from 23.3% to 43.3% with GPT-curated knowledge (Table 5).
Critical weaknesses
  • The main comparisons are not shown to be like-for-like, and knowledge is not separated from repeated attempts "Solved tasks are removed from the active pool" (#4.1, L255), and each later agent receives the attempt table and a bundle for the same task, so a task counts as solved if any of up to 10 attempts succeeds (on ARC, up to 20 scored submissions per test input). Solve rate is not defined for DGM, HyperAgents, GEPA ("best score", Table 3) or OpenEvolve. If these are single-pass scores of a final or best agent or prompt, Tables 2 and 3 compare different metrics; if they are unions over all agents evaluated, the concern narrows to the same-task memory these four baselines lack and to the missing retry controls. The Terminal-Bench 2 baselines were not rerun (L345). No generation-1 (no-knowledge, single-attempt) rate is reported, and the only retry control, a single-seed Polyglot run with GPT-5.4-mini, reaches 62.0% without any knowledge, against the protocol's 72.7% three-seed mean (Tables 4 and 8).
  • Evaluator information may enter per-task knowledge Appendix I states that "no method sees hidden tests or gold answers" (L1486). Yet a stored post in Figure 5 reads "The first concrete mismatch occurred at row 1, col 6: expected 0, but 6 was submitted", which looks like a comparison with the hidden ARC test output. Polyglot solvers see "the last 50 lines of its own test-runner output" from the otherwise hidden tests (L1524), and forum posts may carry this into later generations. SWE-bench Pro's seeded mode "can expose failing tests, selected test files, and a benchmark wrapper command" (L1164–1165), and the paper does not say which mode the main runs used. If answer-revealing content reaches per-task bundles, that channel measures leakage rather than knowledge.
Questions for authors
  1. For each method in Tables 2 and 3, is "solve rate over 10 generations" the union of tasks solved in any generation, or one pass of the final or best agent or prompt? For each method, benchmark and LLM, please report the generation-1 solve rate, the cumulative curve over generations 1–10, and one fresh attempt on all 50 tasks with the final artifact (for your method, the frozen generation-10 knowledge base, including already-solved tasks), over three seeds.
  2. What evaluator-derived information reaches forum agents, the typed attempt table and the bundles?
  • (a) Is the Figure 5 mismatch computed against a training output or the hidden test output?
  • (b) Are Polyglot hidden-test logs stored and carried into later generations?
  • (c) Did the main SWE-bench Pro runs use upstream-strict or seeded mode, and does the benchmark test-script command run the hidden FAIL_TO_PASS tests? If hidden-answer content is stored, please rerun the affected benchmarks storing only binary pass/fail outcomes.
  1. Can you run the "independent agents" and "raw attempts" controls of Appendix N with Haiku 4.5, three seeds and the same 10-generation budget on all five benchmarks, and report paired task-level tests (for example, McNemar over tasks) against the full protocol?

  2. For Table 5, can you add (i) an adapter-only condition (the same extra LLM call and memo schema with an empty or unrelated prior) and (ii) the baselines' final artifacts (DGM and HyperAgents best agent, GEPA and OpenEvolve best prompt, ExpeL rules, ACE playbook) evaluated on the same 20 held-out tasks?

  3. Which ARC-AGI-2 split supplies the 50-task pool? Which donor seed's knowledge base was frozen for Table 5? Is the 76.0% full-protocol run in Table 8 separate from the three Table 4 seeds?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.