# Self-Evolving Agents

16 related articles

Subscribe to updates

arXiv Nov 4, 2026
Reglos splits personal-agent memory into two loops: memory grows with each chat while a coding agent rewrites the harness
Reglos, a personal memory agent, splits improvement across two time scales: each user's memory grows online as conversations arrive, while the harness that decides which memories reach the model is revised offline by a coding agent. A critic from a second model family screens patches for benchmark gaming, and a fixed evaluator accepts them only on users held out from development. With Qwen3.8-27B writing memory and answering, Reglos reaches 61.3% on PersonaMem-v2, 7.5 points above the strongest compared system.
Reglos, a personal memory agent, splits improvement across two time scales: each user's memory grows online as conversations arrive, while the harness that decides which memories reach the model is revised offline by a coding agent. A critic from a second model family screens patches for benchmark gaming, and a fixed evaluator accepts them only on users held out from development. With Qwen3.8-27B writing memory and answering, Reglos reaches 61.3% on PersonaMem-v2, 7.5 points above the strongest compared system.
Sen Yang（First author）
Agent Memory Recursive Self-Improvement Reglos Self-Evolving Agents PersonaMem-v2
arXiv Nov 11, 2026
Agents stop writing every lesson to memory: FGSE tests each update for transfer, falsification and regression first
Long-running LLM agents often write each episode's lesson straight into memory, yet local success is not a reliable sign of reusable, safe knowledge. FGSE turns every candidate lesson into a structured hypothesis with a precondition, behavior change, expected effect and verifier, applies it only in a temporary state, and tests it on target-transfer, falsification and archived-regression probes before committing, narrowing or rejecting it. On verifier-rich benchmarks including WebArena, AppWorld and τ-bench, FGSE with a DeepSeek-V3.1 backbone reaches 39.8 WebArena success rate and 78.8 AppWorld average completion, while cutting falsification failures, regression damage and harmful committed updates to 5.0%, 1.3% and 3.6%.
Long-running LLM agents often write each episode's lesson straight into memory, yet local success is not a reliable sign of reusable, safe knowledge. FGSE turns every candidate lesson into a structured hypothesis with a precondition, behavior change, expected effect and verifier, applies it only in a temporary state, and tests it on target-transfer, falsification and archived-regression probes before committing, narrowing or rejecting it. On verifier-rich benchmarks including WebArena, AppWorld and τ-bench, FGSE with a DeepSeek-V3.1 backbone reaches 39.8 WebArena success rate and 78.8 AppWorld average completion, while cutting falsification failures, regression damage and harmful committed updates to 5.0%, 1.3% and 3.6%.
Runxuan Tang（First author）
Falsification-Gated Self-Evolution LLM Agents Persistent Memory Self-Evolving Agents
arXiv Oct 22, 2026
A self-evolving trading agent rewrites its own prompts, but Steward locks risk rules beyond the type system and lost just 0.3% in a 24.6% sector drawdown
Autonomous investing agents must learn from outcomes without rewriting their own risk rules. Steward separates editable prompts from an invariant core: nine role-specialised agents decide jointly, code grades every trade nightly, and a quarterly reflector proposes only typed prompt edits that a deterministic gate applies reversibly. Over 102 trading days and 1,632 model calls in a paper account, the portfolio lost 0.3% while the semiconductor sector fell 24.6%, and no injected instruction was followed.
Autonomous investing agents must learn from outcomes without rewriting their own risk rules. Steward separates editable prompts from an invariant core: nine role-specialised agents decide jointly, code grades every trade nightly, and a quarterly reflector proposes only typed prompt edits that a deterministic gate applies reversibly. Over 102 trading days and 1,632 model calls in a paper account, the portfolio lost 0.3% while the semiconductor sector fell 24.6%, and no injected instruction was followed.
Nanyang Technological University
Yebo Feng（First author）
Steward Self-Evolving Agents multi-agent systems Autonomous Investing Prompt Injection
arXiv Oct 29, 2026
Image-generation agents stop merely retrying: failure experience passes a counterfactual check before entering the skill library, reaching 86.1 on GenEval2
Agentic image-generation pipelines have largely been read-only generate-verify-rewrite loops that discard repair experience or admit ungated updates into a skill library. VEvo-Harness expands each failed verdict into a hierarchical verification tree to locate root causes, then admits only causally valid skill patches through counterfactual dual-arm image replay, with memory organized into working, experience and pattern tiers. On the Qwen-Image-2512 backend it reaches 86.1 on GenEval2, 14.5 points above GenRouter, while continuing to evolve.
Agentic image-generation pipelines have largely been read-only generate-verify-rewrite loops that discard repair experience or admit ungated updates into a skill library. VEvo-Harness expands each failed verdict into a hierarchical verification tree to locate root causes, then admits only causally valid skill patches through counterfactual dual-arm image replay, with memory organized into working, experience and pattern tiers. On the Qwen-Image-2512 backend it reaches 86.1 on GenEval2, 14.5 points above GenRouter, while continuing to evolve.
Yuan Wang（First author）
VEvo-Harness Agent Harness Self-Evolving Agents Image Generation Image Editing
arXiv Oct 16, 2026
Two humanoids carry one payload through a narrow corridor: coordination becomes executable code, and a failed physical check rewrites roles
When two humanoid robots carry one freely supported rigid payload through constrained corridors, team coordination is written as executable code rather than fixed by task decomposition, with a multimodal foundation model supplying motion primitives, semantic objectives, temporary roles and bounded parameters. A verifier judges each physical transition; rejected ones trigger structured revision from the latest accepted checkpoint, and verified experience is retained for reuse. Across 15 base-task runs with three foundation-model backbones, 8 tasks completed, with every backbone achieving end-to-end success.
When two humanoid robots carry one freely supported rigid payload through constrained corridors, team coordination is written as executable code rather than fixed by task decomposition, with a multimodal foundation model supplying motion primitives, semantic objectives, temporary roles and bounded parameters. A verifier judges each physical transition; rejected ones trigger structured revision from the latest accepted checkpoint, and verified experience is retained for reuse. Across 15 base-task runs with three foundation-model backbones, 8 tasks completed, with every backbone achieving end-to-end success.
Beijing Academy of Artificial Intelligence Peking University Renmin University of China
Zicheng Liu（First author）
CoEvolve Verifier Guided Coordination Evolution Multi-robot coordination Cooperative payload transport Self-Evolving Agents
arXiv Oct 12, 2026
Self-evolving agents stop splitting weights from harness: Evo2 feeds each into the other, resolving 48.3% of SWE-bench Verified
Evo2 turns harness edits and their execution outcomes into training signal for the model that writes them, lifting SWE-bench Verified resolution from the strongest baseline's 41.0% to 48.3% and Polyglot pass@1 from 62.6% to 82.1%.
University of Illinois Urbana-Champaign
Chejian Xu（First author）
Evo2 Weight-Harness Co-Evolution EvoTIDE Self-Evolving Agents Crossover Searching
arXiv Oct 14, 2026
Agents stop waiting for an external teacher: ClawLoop has them write tasks, diagnose failures, and retrain in the same sandbox
ClawLoop closes task synthesis, training, and failure diagnosis into one loop so harness-based agents improve without an external teacher model, beating the one-shot ClawGym pipeline on SkillsBench, Claw-Eval-Live, and ClawGym-Bench.
Meituan (China)Peking University
Hansi Yang（First author）
ClawLoop Task Synthesis Self-Evolving Agents GRPO Self-Distillation
arXiv Oct 12, 2026
Agents stop rewriting behavior: falsifiable environment rules lift reactor-lab completion from 0% to 83%
EvoBelief writes an agent's knowledge of its environment as falsifiable hypotheses, and only rules that survive independent evidence may change predictions or plans, while the language model stays frozen. On ALFWorld without actor-side transition memory it raises success over the matched Frozen World Model by 23.9% with Claude Haiku 4.5 and 31.3% with GPT-5.4; in DiscoveryWorld's Reactor Lab it completes 83.3% of Normal-tier and 58.3% of Challenge-tier instances, versus 0% for each evaluated baseline.
EvoBelief writes an agent's knowledge of its environment as falsifiable hypotheses, and only rules that survive independent evidence may change predictions or plans, while the language model stays frozen. On ALFWorld without actor-side transition memory it raises success over the matched Frozen World Model by 23.9% with Claude Haiku 4.5 and 31.3% with GPT-5.4; in DiscoveryWorld's Reactor Lab it completes 83.3% of Normal-tier and 58.3% of Challenge-tier instances, versus 0% for each evaluated baseline.
Amazon (United States)Tulane University
Janet Wang（First author）
EvoBelief World Model Self-Evolving Agents LLM Agents
arXiv Oct 9, 2026
Confining self-evolution to the runtime harness behind one admission gate: 144 of 7,449 candidates admitted with no harmful change, while the same loops without the gate admitted 309 harmful changes
The work argues that self-evolution of LLM agents should be confined to the runtime harness—instruction text, tool-call logic and typed primitive composition—while weights stay fixed, and gives a dual-loop engine in which a local loop edits one primitive and a global loop replaces primitives from a typed library, both submitting to one admission gate; on a synthetic regulatory-shift benchmark (three families of re-interpretation, three severities, 10 seeds per cell) the gate admitted 144 of 7,449 candidates with none worsening error on held-out history, whereas replacing the gate with the check an unbounded system applies—fewer errors visible in recent traces—let the same loops admit 309 harmful changes and left missed flags above 10% in 49 of 90 runs.
The work argues that self-evolution of LLM agents should be confined to the runtime harness—instruction text, tool-call logic and typed primitive composition—while weights stay fixed, and gives a dual-loop engine in which a local loop edits one primitive and a global loop replaces primitives from a typed library, both submitting to one admission gate; on a synthetic regulatory-shift benchmark (three families of re-interpretation, three severities, 10 seeds per cell) the gate admitted 144 of 7,449 candidates with none worsening error on held-out history, whereas replacing the gate with the check an unbounded system applies—fewer errors visible in recent traces—let the same loops admit 309 harmful changes and left missed flags above 10% in 49 of 90 runs.
Ravil Akhtyamov（First author, Corresponding author）
Self-Evolving Agents Agent Harness Admission Gate LLM Agents Model Risk Management
arXiv Oct 9, 2026
SynCo jointly optimizes synthesizer and reasoner in the same online batch, reaching 78.72% average across eight math reasoning benchmarks
SynCo formulates agentic data synthesis and task solving as a coupled multi-agent reinforcement learning problem, updating a Synthesizer and a Reasoner from the same online interaction batch: the Reasoner receives rollout-level correctness rewards while the Synthesizer receives a task-level reward gated by task quality, answer reliability, and outcome-grounded teachability, with both updates feeding the next synthesis round. Across eight mathematical reasoning benchmarks, SynCo reaches 78.72% average accuracy versus 75.85% for a frozen-Synthesizer variant and 74.96% for the strongest external baseline, ScaleQuest-Math, with gains arising mainly from problems the initial model failed to solve.
SynCo formulates agentic data synthesis and task solving as a coupled multi-agent reinforcement learning problem, updating a Synthesizer and a Reasoner from the same online interaction batch: the Reasoner receives rollout-level correctness rewards while the Synthesizer receives a task-level reward gated by task quality, answer reliability, and outcome-grounded teachability, with both updates feeding the next synthesis round. Across eight mathematical reasoning benchmarks, SynCo reaches 78.72% average accuracy versus 75.85% for a frozen-Synthesizer variant and 74.96% for the strongest external baseline, ScaleQuest-Math, with gains arising mainly from problems the initial model failed to solve.
George Washington University Georgia Institute of Technology University of Southern California
Wei Yang（First author）
SynCo Multi-Agent Reinforcement Learning Agentic Data Synthesis Self-Evolving Agents Group-Relative Policy Optimization