Research map / Multi-agent systems and games / LLM agents

Large Language Model Agents Are Not Always Faithful Self-Evolvers

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Systematically tests whether self-evolving LLM agents causally depend on raw and condensed experience via controlled interventions, finding strong reliance on raw trajectories but weak sensitivity to condensed experience.

Problem

Whether self-evolving LLM agents faithfully and causally use the past experience they retrieve (raw trajectories or condensed insights) to guide their decisions.

Approach

Defines experience faithfulness as causal dependence of behavior on experience and probes it with controlled interventions (empty, shuffle, corrupt, irrelevant, filler) on raw and condensed experience across four self-evolving frameworks, finding an asymmetry where condensed experience is largely ignored.

Key results

• Raw experience perturbations (e.g., Empty, Irrelevant) cause substantial drops on ExpeL/HotpotQA and ALFWorld, while condensed-experience interventions (Corrupt, Irrelevant, Filler) yield near-baseline performance across frameworks.
• In ReasoningBank on WebArena with only condensed experience, removing it lowers success but Corrupt/Irrelevant/Filler cause only negligible degradation or slight gains.
• Integrated Gradients shows condensed experience receives consistently low attribution across layers while the current trajectory dominates later layers.

Surprising findings

• Agents sometimes perform better when condensed experience is perturbed or removed.
• Even semantically empty placeholder tokens in condensed experience yield performance similar to real condensed experience.

Limitations

• Condensed-experience interventions tested mainly on four frameworks and fixed prompts; generalization to other memory designs is not established.
• Integrated Gradients analysis is limited to Qwen3 models on ALFWorld with a specific attribution procedure.
• Multi-hop QA intervention analysis uses only 100 sampled examples per benchmark.

Experimental setup

Datasets: HotpotQA FEVER GPQA-Diamond MMLU-Pro Eng. AIME 2024 Game of 24 ALFWorld WebArena WebShop 2WikiMultiHopQA Musique

Benchmarks: HotpotQA FEVER GPQA-Diamond MMLU-Pro Eng. AIME 2024 Game of 24 ALFWorld WebArena WebShop 2WikiMultiHopQA Musique

Metrics: success rate exact match

Baselines: ExpeL Dynamic Cheatsheet ReasoningBank G-Memory

Builds on: ExpeL Dynamic Cheatsheet ReasoningBank G-Memory

Closest papers by meaning

• MemEvolve: Meta-Evolution of Agent Memory Systems (ICML 2026 · similarity 0.90)
• EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective (NeurIPS 2026 · similarity 0.89)
• Bias Amplification in Language Model Evolution: An Iterated Learning Perspective (NeurIPS 2024 · similarity 0.88)
• Can large language models explore in-context? (NeurIPS 2024 · similarity 0.88)
• PACE: Two-Timescale Self-Evolution for Small Language Model Agents (NeurIPS 2026 · similarity 0.88)
• Reflection-Bench: Evaluating Epistemic Agency in Large Language Models (ICML 2025 · similarity 0.88)