6arXiv cs.CL (Computation and Language)·
Continual-learning evaluation on Terminal-Bench 2.0 tests whether agent optimizer gains compound across tasks
A new arXiv paper introduces a two-phase continual-learning evaluation framework built on Terminal-Bench 2.0 to test whether agent-optimization gains persist and compound when new tasks arrive over time. Three agent-harness optimization methods — GEPA, Meta Harness, and RELAI-VCL — are compared under identical budgets; all improve in static single-phase settings but diverge sharply under continual optimization. RELAI-VCL is the only method that both transfers positively to unseen tasks and continues improving, reaching a 76.4% lifelong average pass rate versus 58.7% for the unoptimized baseline. The key finding is that compounding gains require regression control built into the optimization loop to prevent shortcut solutions.
Related guides (3)
Related events (8)
AgentCL: A Rigorous Evaluation Framework for Continual Learning in Language Agents
AgentCL is a new benchmark and evaluation framework designed to rigorously assess continual learning in language agents, addressing gaps in existing benchmarks that focus on retrieval over long-context documents or use naive task streams with limited cross-task analysis. The framework constructs compositional task streams where earlier sub-solutions, evidence, or workflows are intentionally reusable in later tasks, contrasting them with naive streams to measure transfer gains. The authors also introduce MemProbe, a probing method that stores interactions, insights, and skills while filtering unreliable experiences during consolidation. Empirical results across coding, deep research, and language understanding tasks show that controlled streams better distinguish memory design quality, and that naive streams can mask memory-induced degradation.
Self-evolving agent harness framework uses quality-diversity archive to generalize harness improvements across model families
A new arXiv preprint introduces a framework for automatically improving LLM agent harnesses (prompts, injected knowledge, runtime control, configuration) without modifying model weights. The system separates change proposal (done by an LLM diagnosing failures) from credit assignment (done by deterministic statistical testing), populating a gated quality-diversity archive keyed on failure pathology rather than task identity to prevent overfitting. Evaluated across seven domains with a frozen open-weight model, the framework achieves +9 to +15.5 percentage point gains on sealed test sets, retaining 86–147% of training gains, with the key finding that pathology-to-patch mappings transfer across model families.
EvoPolicyGym benchmark evaluates autonomous policy evolution by LLM agents in RL environments
A new arXiv preprint introduces EvoPolicyGym, a benchmark for evaluating how LLM-based agents iteratively improve executable policies in compact interactive RL environments under a fixed interaction budget. The benchmark provides trajectory-level diagnostics beyond aggregate scores, distinguishing how agents allocate budget and convert feedback into parametric tuning. GPT-5.5 achieves the strongest aggregate rank score and top-two performance across all 16 environments. The work targets a gap in agent evaluation where iterative policy refinement is conflated with open-ended software engineering progress.
Double Ratchet: Co-evolving evaluation metrics and skills for self-improving LLM agents
A new arXiv preprint introduces Double Ratchet, a system that co-evolves both evaluation metrics and agent skills in settings where no reliable automatic verifier exists. The metric loop uses evolutionary search over small drawback detectors anchored to a small reference set, while the skill loop uses a lifecycle-managed approach; together they retain 88–110% of the performance lift achievable with ground-truth metrics across code generation (MBPP+), text-to-SQL (Spider 2.0-Snow), and report generation tasks. The paper also addresses safety, showing that anchor discipline and outer audits can catch and repair cases where evolved skills game the rubric. This work directly addresses a core bottleneck in self-improving agent systems: the chicken-and-egg problem of needing a reliable evaluator to improve.
Tmax: Open RL training recipe for terminal-using agents achieves 27% on Terminal-Bench 2.0 with 9B parameters
Researchers present Tmax, an open RL training recipe for terminal-using language model agents, achieving 27% on Terminal-Bench 2.0 with a 9B parameter model while outperforming larger models from prior work. The recipe combines a novel data generation taxonomy using difficulty control, personas, and verifier diversification to produce a terminal environment dataset over 2.5x larger than previously released datasets. Training uses a simple outcome-only RL approach, and the authors release data, models, and code to lower the barrier for academic research on terminal agents.
RePro: Retrospective Progress-Aware Self-Refinement for LLM Agent Training
Researchers introduce RePro (Retrospective Progress-Aware Training), a framework addressing the gap between step-wise RL optimization and metacognitive task-progress awareness in LLM agents. The approach uses a forward-then-reflect rollout paradigm where agents execute actions online and then retrospectively assess step-wise progress given the completed trajectory and known outcome. Evaluated on WebShop, ALFWorld, and Sokoban, RePro achieves up to 12% absolute success rate gains over baseline Qwen-family models without requiring continuous external supervision.
AutoLab benchmark evaluates frontier models on ultra long-horizon iterative research and engineering tasks
AutoLab is a new benchmark of 36 expert-curated tasks across system optimization, puzzle-solving, model development, and CUDA kernel optimization, designed to test agents on sustained closed-loop improvement under wall-clock budgets rather than single-turn or short-horizon settings. Evaluation of 17 frontier models finds that persistence in iterative benchmarking and feedback incorporation — not initial attempt quality — is the dominant success predictor. Claude Opus 4.6 stands out as the strongest performer, while most models including proprietary ones either terminate early or exhaust budgets with minimal progress. The benchmark, harness, and task artifacts are open-sourced.
EdgeBench finds log-sigmoid scaling laws for agent learning from real-world environments
A new arXiv preprint introduces EdgeBench, a suite of 134 real-world tasks with ultra-long horizons (12+ hours of continuous agent operation each) spanning scientific discovery, software engineering, formal mathematics, and other domains. Analyzing ~38,000 hours of agent-environment interaction, the authors report the first evidence that agent performance during environment learning follows a log-sigmoid scaling law with R²=0.998. They also find that agent learning speed roughly doubles every three months across model generations, drawing an analogy to pretraining scaling laws but for post-deployment environment learning. 51 tasks and the evaluation framework are publicly released.