Source-linked AI summary
Evo$^2$: Evolving the Evolver through Weight-Harness Co-Evolution
Chejian Xu, Shujian Zhang, Wenxuan Zhou, John Lambert, Beck Chen, Bo Li, Andrew Hard, Rajiv Mathews
TL;DR
Existing self-evolution approaches typically improve model weights or agent harnesses separately, despite their relationship. Evo2 couples harness evolution with training a shared Proposer–Reflector policy, and reports improvements across coding benchmarks alongside harness transfer across benchmarks and tasks.
Problem
Self-evolution methods typically update model weights or modify agent harnesses separately, although model capabilities shape harness edits and evaluated edits can inform later weight updates.
Source: §1 · pp. 1–2Method
Evo2 uses one policy as both Proposer and Reflector, trains it on harness edits and execution outcomes, and uses EvoTIDE to recover trajectories for SFT and interleaved GRPO.
Source: §1 · p. 2Results
Across three coding benchmarks, Evo2 outperforms existing self-evolution methods, achieving 82.1% on Polyglot versus 62.6% for the strongest baseline and 48.3% on SWE-bench Verified versus 41.0%.
Source: §1 · p. 2Takeaways & Limitations
Evolved harnesses transfer across benchmarks, and the dual-role harness-improvement framework also applies to fixed-weight proprietary models and memory tasks.
Source: §1 · pp. 2–3, §5 · p. 11
Abstract
from arXiv · showhide
Self-evolving agents are increasingly used to improve language model systems without human intervention through either weight updates or harness improvements. Existing approaches typically pursue these directions separately: some methods update model weights using self-generated experience, while others rewrite the harness that defines how an agent uses prompts, tools, context, and execution feedback. However, these approaches overlook the relationship between model weights and the agent harness. To connect these processes and ensure effective and smooth agent self-evolving, we propose Evo$^2$, a framework for weight-harness co-evolution. For harness updates, we design a unified dual-role pipeline that a single policy serves as both a Proposer, which writes harness edits, and a Reflector, which diagnoses their effects and guides revision. For weight updates, harness modifications and their execution outcomes guide the training of both the Proposer and the Reflector. We further develop EvoTIDE, a training pipeline that recovers usable trajectories through edit extraction and prefix continuation, then combines SFT with interleaved GRPO across the two roles. Extensive experiments across three coding benchmarks demonstrate that Evo$^2$ significantly outperforms existing self-evolution methods. It achieves 82.1% on Polyglot and 48.3% on SWE-bench Verified, compared with 62.6% and 41.0% for the strongest baseline, respectively. We also show that the evolved harnesses transfer to different tasks without further evolution and rank first in cross-benchmark evaluations. Our dual-role harness improvement framework is also able to extend to self-evolution with proprietary models, improving a GPT-5-mini agent's performance from 24.4% to 48.3%. These results demonstrate the applicability of the framework Evo$^2$ across domains and its broad transferability across models and tasks.
1. Introduction
Existing self-evolution methods typically improve model weights or agent harnesses separately, although harness edits and their outcomes can also teach the model that guides later edits. Evo2 connects these processes through shared Proposer–Reflector training and reports stronger coding results and cross-benchmark harness transfer.
1. Introduction
- Weight updates and harness edits are usually pursued separately, despite their reciprocal relationship: model capabilities shape edits, and evaluated edits reveal useful changes and failures.The paper frames proposing edits and diagnosing their effects as skills learnable from earlier evolution attempts.Source: §1 · pp. 1–2
- Evo2 trains a shared policy from harness modifications and execution outcomes, so updated weights guide subsequent harness search.The policy serves as both Proposer, which writes edits, and Reflector, which interprets outcomes and guides revision.Source: §1 · p. 2
- Crossover searching lets the Reflector select a useful strategy from one branch and the Proposer adapt it to another harness.The framework combines this search mechanism with EvoTIDE, which recovers usable trajectories and trains both roles using SFT followed by interleaved GRPO.Source: §1 · pp. 2–3
- 82.1% on Polyglot and 48.3% on SWE-bench Verified exceed the strongest respective baselines, 62.6% and 41.0%.The evolved harnesses also transfer across benchmarks without further evolution, and the framework extends to memory tasks and fixed-weight proprietary models.Source: §1 · pp. 2–3
2. Related Work
Prior self-improvement research trains models from generated experience, while self-evolving-agent work modifies harnesses and evaluates the resulting agents. Evo2 connects these lines by using harness-evolution experience to train a shared policy.
2. Related Work
- Self-improvement methods train from model outputs, self-play, selected trajectories, or learned feedback mechanisms, with some methods jointly training generation and feedback.Examples include STaR and ReSTEM [Singh et al., 2024], and EvoLM and CoNL [Li et al., 2026; Sui and Hooi, 2026].Source: §2 · p. 3
- Harness-evolution methods edit agent code, prompts, or control flow and use evaluations or execution traces to guide later changes.Prior work also studies branching search, reflection, shared experience, and transfer across tasks or models, while Evo2 trains its evolver on evaluated edits and outcomes.Source: §2 · p. 3
3. Method
Evo2 alternates harness search with training of a shared Proposer–Reflector policy, using candidate evaluations and role trajectories as experience. EvoTIDE recovers and trains on those trajectories so updated weights can guide later search.
3.1. Overview
- The fixed task executor evaluates harness candidates, isolating the effect of harness changes while the policy parameters are trained on modification outcomes.Search samples a parent, proposes and evaluates edits, then uses reflection to recommend revision, crossover, or retention; resulting candidates extend the tree.Source: §3.1 · p. 4
3.2. Harness evolution through dual-role self-modification
- The shared policy proposes autonomous or directed edits, while the Reflector diagnoses execution outcomes and chooses revision, crossover, or keep.Crossover selects a strategy from an evaluated branch for adaptation to the current candidate; the resulting harness is evaluated and remains a child of the recipient.Source: §3.2 · pp. 4–5
3.3. Weight evolution through EvoTIDE
- EvoTIDE collects trajectories from both roles, first uses supervised fine-tuning, then trains both roles with interleaved GRPO.The updated policy generates later proposals and feedback.Source: §3.1 · p. 4, §3.3 · p. 5
3.3.1. Data collection
- Trajectory recovery extracts concrete file edits from rollouts lacking explicit edit actions and continues valid prefixes after later edits fail.Recovered proposals still pass application, startup, and task-evaluation checks; reflection cases include improvements, unchanged results, regressions, and crossover strategies.Source: §3.3.1 · pp. 5–6
3.3.2. Reward design
- Valid Proposer and Reflector actions receive execution rewards based on task-score changes, while invalid or unusable edits are penalized.A keep action has zero execution reward; Reflector gains count only when traces show the added mechanism ran.Source: §3.3.2 · p. 6
3.3.3. Training process
- SFT initializes the policy from successful autonomous edits, directed edits, and Reflector responses selected by downstream evaluation.Training balances roles and modes, then interleaves Proposer and Reflector groups; a frozen snapshot supplies counterpart behavior during each round.Source: §3.3.3 · pp. 6–7
- The parent pool retains harnesses at multiple depths and includes both strong candidates for further improvement and weaker candidates for repair.The trained evolver can search from a new initial harness, while evolved harnesses can be deployed on new tasks with compatible model and tool interfaces.Source: §3.3.3 · pp. 6–7
4. Experiments
Across coding benchmarks, cross-benchmark transfer, proprietary models, and memory tasks, Evo2’s evolved harnesses outperform baselines. Analyses also examine weight evolution and inference cost.
4.1. Experimental Settings
- Experiments use Qwen3.8-27B as executor and evolver for coding, evaluate Polyglot, SWE-bench Verified, and LiveCodeBench, and use fixed-weight GPT-5-mini and Claude Haiku 4.5 in proprietary-model tests.Memory experiments evaluate LoCoMo and LongMemEval; coding reports held-out pass@1 on Polyglot and LiveCodeBench and resolved rate on SWE-bench Verified.Source: §4.1 · p. 8
4.2. Self-Evolution on Coding Benchmarks
- 82.1% pass@1 on Polyglot and 48.3% resolved rate on SWE-bench Verified exceed the strongest respective baselines, 62.6% and 41.0%; Evo2 reaches 58.0% pass@1 on LiveCodeBench.On LiveCodeBench, the four baselines range from 39.0% to 56.0%.Source: §4.2 · p. 8
4.3. Harness Transfer across Benchmarks
- Evo2 achieves the best transfer in five of six directions; its SWE-bench-evolved harness scores 72.4 on Polyglot and 53.0 on LiveCodeBench.Table 2 evaluates each benchmark-selected harness on the other two benchmarks without further evolution.Source: §4.3 · pp. 8–9, Table 2 · p. 9
4.4. Self-Evolution with Proprietary Models
- 48.3% resolved rate with GPT-5-mini exceeds baseline rates of 19.4–24.4%; with Claude Haiku 4.5, Evo2 reaches 47.8% versus 42.2–45.0%.The models serve as executor, Proposer, and Reflector with fixed weights.Source: §4.4 · p. 9, Table 3 · p. 9
4.5. Self-Evolution on Memory Tasks
- On LoCoMo, Evo2 reaches 60.4 F1 versus HGM’s 56.2; on unseen LongMemEval, its searches score 67.9–73.9 versus HGM’s 64.4 mean.The same harness is evaluated on LongMemEval without further evolution or selection.Source: Table 4 · p. 9, §4.5 · p. 10
4.6. In-Depth Analysis of Evo2
- With search, executor, and parent harnesses fixed, EvoTIDE-trained weights raise champion performance by +16.3 points on Polyglot, +9.3 on LoCoMo, and +10.6 on LongMemEval transfer.Crossover search adds a further 4.1-point improvement on Polyglot.Source: §4.6 · p. 10, Table 5 · p. 10
- Evo2’s LiveCodeBench harness uses 3.6K completion tokens and 10.2 calls per task; HGM and DGM use 6.6K–6.9K tokens and 26–37 calls despite lower accuracy.On Polyglot, Evo2’s champion uses 36 calls and 3.6K tokens per task, versus 43 calls and 4.0K tokens for the unevolved harness.Source: §4.6 · p. 10, Figure 3 · p. 10
5. Conclusion
Evo2 co-evolves agent harnesses and the policy that guides their evolution. Experiments on coding and memory tasks show performance gains and transfer, including with fixed proprietary models.
- A shared policy acts as Proposer and Reflector, learning from harness edits and execution outcomes to propose changes and guide revisions.The framework combines harness evolution with weight evolution.Source: §5 · p. 11
- Crossover search combines strategies across harness branches, while EvoTIDE trains the shared policy through trajectory recovery, SFT, and interleaved GRPO.Source: §5 · p. 11
- Experiments support improved performance and harness transfer across coding and memory tasks, and dual-role search also improves agents with fixed proprietary-model weights.The analyses report improved search reliability from weight evolution and reduced evaluation budget from crossover searching.Source: §5 · p. 11