AI Coding Live News

Microsoft Research and Stanford University have unveiled AutoEnvScaling, an automated data flywheel that frames RL environment creation itself as a terminal execution task. A proposer agent builds diverse environments directly from solver rollouts, slashing cost per valid environment by 75% while achieving a 25x higher valid environment yield over prompt-only baselines. Crucially, unifying environment design and solving within a single model enables Recursive Self-Improvement (RSI), yielding +12.2 points on Terminal-Bench 2.1 and doubling scores on Terminal-Bench 4.0.

Zoom

While LLM agents are rapidly deployed into sandbox workspaces with direct shell and filesystem access, post-training methods remain predominantly one-shot and bound by external teacher models. ClawLoop introduces a self-improving, harness-native training loop that operates without external teachers. Through a self-sustaining cycle of task synthesis with difficulty growth, SFT/RL training, and automated behavioral diagnosis, ClawLoop demonstrates consistent empirical gains across SkillsBench, Claw-Eval-Live, and ClawGym-Bench on Qwen3.6-27B, while successfully transferring data to 4B and 9B models.

Zoom

On Oct 14 OpenAI's API deprecations page announced that GPT-4o (2024-08-06 and 2024-11-20 snapshots), GPT-4o mini (2024-07-18), GPT-4.1 and GPT-4.1 mini (2025-04-14) will be removed from the API on January 14, 2027. Fine-tunes built on them shut down the same day, and the recommended replacement for all four is gpt-6-luna, listed at $0.10 input / $0.50 output per 1M tokens versus $2.50 / $10 for GPT-4o.

Zoom

On Oct 13 Cursor launched the Origin App Marketplace: its code hosting platform Origin now connects to 20+ apps across deployments (Cloudflare, Vercel and more), CI (CircleCI, Buildkite and more), code review (CodeRabbit, Greptile, Tenki), security (Semgrep, Socket and more), monitoring (Sentry) and project management (Linear), and the marketplace is open to public submissions from any developer-tool team.

Zoom

Researchers from Shanghai Jiao Tong University have introduced Multi-Stream Attention Residuals (MSAR), extending Attention Residuals (AttnRes) to multi-head attention along the network depth axis. By integrating multi-scale causal convolutions to retrieve distinct contextual views and combining them via adaptive mixing, MSAR achieves multi-stream capacity without persistent communication or memory overhead. Evaluated across 18B and 28B MoE scales, MSAR delivers a 1.58x iso-loss compute speedup over standard residual baselines.

Zoom

Salesforce AI Research has introduced Prospective Hindsight (PH), a self-calibrating training framework for long-horizon autonomous agents. Addressing the fundamental limitation of standard RL where credit assignment occurs strictly after environmental feedback, PH exploits the gap between the agent's pre-action prospective prediction and post-action retrospective evaluation. By weighting gradients with this surprise residual, PH eliminates overconfident failures in single-turn tasks and underconfident successes in multi-turn interactions across GRPO and on-policy distillation.

Zoom

Robotic manipulation and physical intelligence systems suffer when relying on general-purpose vision backbones that struggle with dense clutter, tiny objects, and closed-loop control speeds. GroundingPI establishes a dedicated 4B grounding foundation model that unifies points and bounding boxes as quantized coordinates within a shared vocabulary. Outperforming 44 baselines across 34 benchmarks with an average of 73.68% (surpassing GPT-6 Astra at 71.54%), GroundingPI sets a new standard for embodied robotic visual backbones.

Zoom

On-policy distillation (OPD) is critical for reasoning and agent alignment, yet conventional methods realize only a single continuation per decision, severely restricting supervision under finite compute budgets. Researchers have introduced BranchOPD (Branching On-Policy Distillation), an algorithm that dynamically identifies high-uncertainty and teacher-student disagreement decision nodes to branch and resample continuations from shared prefixes. BranchOPD slashes generation costs by 19%–31% on agent benchmarks while boosting math reasoning accuracy by 18.55 percentage points.

Zoom

Reinforcement learning for language model agents is constrained by sparse episodic rewards and inefficient surface token alignment. Researchers have introduced PR-OPD (Privileged Representation On-policy Self-Distillation), discovering that injecting hindsight skills alters less than 25% of token probabilities while reconfiguring over 80% of internal hidden states. By aligning multi-layer projected representations against a stop-gradient self-teacher, PR-OPD boosts WebShop accuracy by 14.0 points and ALFWorld by 4.7 points with zero inference overhead.

Zoom

Researchers from Stanford University and the University of Oxford have published an empirical study examining 'no-thinking' behavior in reasoning LLMs. While developers frequently prompt models to disable explicit thinking to reduce latency and token consumption, the authors discover pervasive 'Thinking Inertia': models persist in generating explicit inference traces even under strict negation instructions, revealing an inherent trade-off between answer-only compliance and task accuracy.

Zoom

On Oct 13 Physical Intelligence opened a beta of the Pi API: teams upload robot demonstrations, fine-tune Pi's latest robot foundation model on Pi-managed compute, and serve the result to their robots through a hosted inference gateway. Showcased deployments range from an operator-less wheel loader to strawberry harvesting and server assembly; early-access customer Almond got a deburring policy for its Axol dual-arm robot after four days of iteration. Access is by application and pricing/API parameters are not published; self-hosters can still use the Apache-2.0 openpi repo (π0/π0.5 base checkpoints).

Zoom

Arena released an agentic version of its Max router. Trained on 2.5M+ real Agent Arena sessions, it picks between Fable 5.1 (Max), Astra (Max) and Astra (Medium) per task based on complexity, task type and cost. In a live test it ranked #1 in Agent Arena with a +14.7% score at a $3.27 median cost per task — 0.3% better than the top public model, Fable 5.1 (Max), at 37.5% lower cost — and sent 67% of coding sessions to the cheaper Astra (Medium). It is free to try in Agent Mode for a limited time; API access is via a signup form.

Zoom