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

PACE: Two-Timescale Self-Evolution for Small Language Model Agents

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 0 citations (OpenAlex)

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.

Proposes PACE, a two-timescale self-evolution framework where frozen small LMs (4B–14B) first refine prompts and only after prompt gains saturate propose control-logic (solver) edits that are committed only if they pass a held-out validation gate.

Problem

Self-evolution of agents usually assumes frontier models for diagnosing failures, proposing revisions and judging updates; it is unclear whether frozen small language models can evolve their own prompts and control logic reliably under resource constraints.

Approach

Schedule low-risk prompt evolution as the default adaptation tool and invoke higher-risk control-logic (solver code) edits only once prompt gains fall below a threshold, separating SLM proposals from acceptance via a validation gate (accept only if the candidate beats the current agent by a margin within budget).

Key results

• PACE achieves the best score on all 12 backbone-benchmark combinations, with up to +9.2% relative gain over vanilla (Qwen3.5-9B IFEval, 0.761 vs 0.697).
• On Qwen3.5-9B MMLU, PACE reaches 0.889 vs 0.835 for +CE and 0.832 for +PE, approaching DeepSeek-V3.2 at 0.908.
• On tau-bench Retail with Qwen3.5-9B, PACE raises pass^1 from 0.791 to 0.817 over vanilla while +PE slightly degrades it to 0.785.

Surprising findings

• Prompt-only evolution degrades tau-bench results on Qwen3.5-9B Retail and collapses on Qwen3-4B Airline (pass^4 0.040 vs 0.120 vanilla).
• Structural (control-logic) evolution from the start is noisy; disabling the validation gate degrades IFEval by about 3.8%.

Limitations

• Evaluation uses small controlled subsets (200 train, 100-800 test) rather than full benchmarks, and results are described as within-protocol comparisons.
• Structural edits can be noisy; proposal quality is weaker for the 4B model, which yields fewer executable and accepted proposals.
• Accepted control-logic edits increase inference cost (up to about 5.1x tokens per query).

Experimental setup

Datasets: MMLU MGSM HotpotQA IFEval tau-bench

Benchmarks: MMLU MGSM HotpotQA IFEval tau-bench

Metrics: accuracy exact match strict instruction accuracy pass^k

Baselines: Vanilla SLM +PE +CE GEPA MIPROv2 Gödel Agent ACE DeepSeek-V3.2

Builds on: GEPA ACE MIPROv2 Gödel Agent

Closest papers by meaning

• MemEvolve: Meta-Evolution of Agent Memory Systems (ICML 2026 · similarity 0.90)
• MAS$^2$: Self-Generative, Self-Configuring, Self-Rectifying Multi-Agent Systems (ICLR 2026 · similarity 0.90)
• Gaia2: Benchmarking LLM Agents on Dynamic and Asynchronous Environments (ICLR 2026 · similarity 0.89)
• EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective (NeurIPS 2026 · similarity 0.89)
• AdaMEM: Test-Time Adaptive Memory for Language Agents (ICML 2026 · similarity 0.89)
• SkillOpt: Executive Strategy for Self-Evolving Agent Skills (NeurIPS 2026 · similarity 0.89)