AI Coding Live News

An LLM agent's practical capabilities depend as much on its execution harness as on its underlying model. Deploying recursive self-improvement (RSI) in open worlds is hampered by the absence of ground-truth verifiers and poor generalization across novel domains. Google Cloud AI Research introduced Harness², a two-level recursive framework: task-level recursion refines task-specific harnesses via a contrastive Propose-Probe-Compose step without ground-truth verifiers, while harnessing-level recursion extracts continual adaptation experience. Harness² delivers up to +11.8% on LAB, +11.5 points on WorkBuddy-Bench, and +4.5 points on JobBench on GLM-5 and Gemini 3.5 Flash. Code is open-sourced on GitHub under Apache-2.0 (google-research/harness2), supporting OpenCode and Codex executors plus JobBench, LAB and WorkBuddy-Bench.

Zoom

When model capabilities exceed human expert evaluative bandwidth, achieving Recursive Self-Improvement (RSI) without external verifiers or compilers becomes the paramount frontier of AGI. While reinforcement learning with verifiable rewards (RLVR) thrives in formal code and math, open-ended research and complex reasoning tasks lack scalar ground truth, and solitary models evaluating their own outputs suffer from homogeneous confirmation bias. Sakana AI introduces MASS (arXiv:2610.12176, project: sakana.ai/mass), an RSI framework anchored in Multi-Agent Self-Supervision. Driven by a single base model (such as Qwen3.6-27B), MASS tasks the model with concurrently acting as task solver, workflow optimizer, and self-evaluator. The framework alternates between evolutionary multi-agent workflow discovery and supervised fine-tuning (SFT) over synthesized trajectories. Across two successive MASS self-evolution cycles, performance per output token increases by 1.2x to 1.6x across open-ended research suites (MLR-Bench, ScienceAgentBench, DSBench, AstaBench). Head-to-head win rates against the base model surge from 53.9% in round 1 to 69.9% in round 2, while a student model trained on multi-agent traces decisively beats a single-agent baseline trained on 1.4x more tokens.

Zoom

As autonomous LLM agents navigate complex long-horizon operational streams in production, enabling continuous test-time weight adaptation without catastrophic policy drift remains a foundational challenge. Conventional in-context adaptation stores reflections or skills as text, making reuse bottlenecked by fragile retrieval heuristics over a static, frozen policy; conversely, directly imitating or reinforcing tokens from single deployment rollouts destabilizes policy weights. Researchers formalize Online Agentic Test-Time Training (OaTTT) and introduce ASCENT (arXiv:2610.05303). The agent executes each task in a single streaming pass. A frozen copy of the initial model acts as a hindsight teacher, receiving verified execution trajectories as privileged context to compute calibrated next-token predictive distributions. Distilling these distributions into persistent LoRA fast weights continuously updates policy parameters across subsequent tasks without external teacher models or gold references. By pruning invalid-action turns to distill condensed execution trajectories, ASCENT systematically improves task success rates and interaction efficiency across ALFWorld, WebShop, and AppWorld, successfully transferring learned behaviors to held-out environments.

Zoom

Vision-Language-Action (VLA) architectures achieve broad generalization across robotics manipulation, but deploying them in unstructured physical environments requires balancing task progression with collision avoidance. Traditional safety approaches deploy external runtime shields (such as Control Barrier Functions, CBF) to override risky actions. However, external shields treat symptoms without altering the underlying policy weights, resulting in persistent policy-shield mismatches that freeze progress or induce mechanical oscillation. Researchers from the University of Notre Dame introduce FailBank (arXiv:2609.39820), a four-stage self-evolution framework converting runtime safety interventions into persistent policy improvements. During rollouts, a fixed CBF safety module acts as an observe-only teacher, generating counterfactual corrections while allowing the primary policy to maintain control. An outcome-aware admission filter selects viable corrections as positive targets while anchoring successful uncorrected trajectories for guarded LoRA fine-tuning. Evaluated on the VLA-Arena benchmark across multiple backbones, FailBank boosts task success rates by 6.9 to 8.5 percentage points while reducing collision costs by 23.8% to 35.6%. Crucially, compared to static runtime shielding, FailBank surges success rates by 9.5 to 25.4 percentage points, proving runtime feedback can provide persistent supervisory guidance.

Zoom

Recent attempts to scale tool-use post-training for autonomous agents have focused primarily on synthesizing standalone execution environments. However, environments constitute merely one component of a holistic agentic interaction system comprising environments, tasks, harnesses, and evaluators; scaling environments in isolation yields inconsistent training signals. Researchers from Fudan University introduce WEFT (arXiv:2609.36887), a whole-system evolution framework for tool-use post-training. WEFT coordinates environment breadth, task complexity, and interaction diversity, applying execution-driven self-evolution to attribute failures and iteratively revise flawed components. To ensure optimization stability and concurrency reliability, WEFT introduces prefix-preserving sampling, atomic-turn credit assignment, and MegaMCP—a Model Context Protocol infrastructure providing isolated, recoverable state management across concurrent rollouts. WEFT-8B and WEFT-14B outperform matched environment-scaling baselines across BFCL V4, tau^2-Bench, and Claw-Eval, with WEFT-14B beating Agent-World-14B by 6.41, 2.23, and 12.27 percentage points, while WEFT-35B-A3B excels on long-horizon benchmarks like Toolathlon-Verified.

Zoom

Self-evolving search agents construct synthetic training curricula by jointly co-optimizing a question proposer and an answer solver. However, this closed loop frequently triggers a deceptive failure mode termed 'co-cheating': the proposer and solver increasingly converge on shared hallucinations, artificially driving up internal training rewards while external factual correctness degrades. The acclaimed research 'False Frontiers' (arXiv:2609.39102, 330+ upvotes on Hugging Face) diagnoses this structural blind spot and introduces CrossFit. By partitioning source documents into orthogonal subsets and scoring proposals with solvers trained exclusively on complementary splits, CrossFit severs the circular pseudo-label reinforcement loop. Evaluated across seven downstream search benchmarks, CrossFit elevates average scores by 8.8 and 8.4 points on 4B and 9B models, systematically outperforming Search-R1.

Zoom

EverMind-AI/Raven orchestrates built-in Research/Code/Design/Oncall plus third-party harnesses (Claude Code, Codex, Hermes, Pi, …) via a Host Agent DAG with admission checks, and frames controlled self-evolution of Raven’s own harness (RSI). Verified: ★5097 / 145 forks; v0.2.3 (2026-09-27); arXiv 2609.33439. Site: Raven-Research 76.5% DeepResearch Mixed; paper: MAOB Exact Match +10.4/+10.5 pp vs strongest baseline. Pre-alpha.

Zoom

Self-evolving search agents jointly optimize a question proposer and an answer solver to generate synthetic training curricula. However, False Frontiers exposes a critical failure mode: 'co-cheating', where proposer and solver agree on shared hallucinations and erroneous pseudo-labels, causing internal rewards to surge while real external accuracy collapses. The authors introduce CrossFit, which partitions source corpora into disjoint splits and scores proposals via cross-trained solvers. Evaluated on Qwen3.5-4B/9B across 7 downstream search benchmarks, CrossFit suppresses false agreement down to 3.7% and outperforms coupled self-evolution by 8.8 points and Search-R1 by 8.7 points.

Zoom

Researchers from UIC exposed two structural flaws in autonomous agents reliant on human-curated toolkits: expert-selected tools frequently degrade anomaly detection across all backbones, while unguided self-revision silently breaks 38% of previously correct responses. They introduce TimeEvo, an architecture that clusters runtime failures into capability gaps, synthesizes evidence-only Python tools to address them, and admits candidate tools via paired verification gates. Starting from an empty toolkit, TimeEvo achieves consistent accuracy gains across all ten benchmarks.

Zoom