Latest news, beside the leaderboard

While deep research (DR) agents excel on synthetic and academic queries, their reliability in realistic personal life and professional workflows remains uncharted. Tencent Hunyuan introduces LifeSearch, a rigorous benchmark featuring 203 real-user tasks across 7 domains backed by 4,500+ hours of expert annotation and 4,000+ fine-grained rubrics. Evaluating 12 leading DR agents reveals that the strongest frontier agent scores only 63.4%, unmasking persistent bottlenecks in multi-step information synthesis and constraint verification.

Vision-language navigation requires abundant, high-quality language supervision, yet traditional synthetic pipelines generate supervision from static models that lag behind evolving navigation policies. Shanghai AI Lab, Tsinghua, and SJTU introduce RSI-Nav, an embodied recursive self-improvement framework combining instruction-following and instruction-generation in a single model. Using shard-wise dual bootstrapping on unlabeled trajectories, RSI-Nav achieves state-of-the-art 68.6% SR on R2R-CE and 72.4% SR on RxR-CE without human intervention.

Off-policy data in reinforcement learning with verifiable rewards (RLVR) causes severe policy gradient bias during asynchronous generation and response reuse. Researchers from UIUC and HKUST introduce Proximal KL, theoretically proving that full future KL credit renders the exact proximal optimizer the unique stationary policy without requiring full-response weights or fitted critics. Evaluated across Qwen3-1.7B and 4B models, Proximal KL outperforms GRPO on 8 math benchmarks by up to 2.58% while completely averting training collapse.

Robotics research typically separates world models (predicting scene transitions) from control policies (generating actions), fragmented by proprietary robot-specific kinematic formats. Sber AI researchers introduced GreenWAM, a foundational world model for robotics. GreenWAM unifies 20 robot embodiment families (from bimanual arms to dexterous humanoids) under a common joint coordinate space, co-training across visual, action-conditioned, and world-action objectives with specialized action-expert heads. GreenWAM achieves top-tier 3D trajectory and depth accuracy on WorldArena 1.0, establishes state-of-the-art policy benchmark scores on RoboTwin 2.0 and LIBERO-plus, and successfully drives real-world humanoid robots outperforming Cosmos 3 Edge and pi_0.5.

In production agentic loops, fast single-pass 'System 1' models are favored for latency-sensitive tool selection and interface actions. When these models fail, engineers cannot easily diagnose whether the agent misperceived environmental observations or misjudged its subsequent action. Researchers from Carnegie Mellon University (CMU) introduced a behavioral diagnostic benchmark across 17 fast models and frontier reasoning baselines, pairing next-action queries with independent factual probes about the state. The results expose a striking systemic pattern: the vast majority of erroneous actions follow completely accurate perception, with error concentration increasing in more accurate models. The true failure mode lies not in reading facts, but in synthesizing them into the next programmatic sub-goal.

Pixel-space world models for GUI agents require predicting future screen images prior to action selection. However, full-screen pixel reconstruction suffers from extreme latency and frequently triggers identity shortcuts, where generated screens merely copy the past rather than reflecting action-induced mutations. Researchers from KAIST introduced the first Latent World Action Model for GUI agents, shifting predictive objectives from pixel generation to action-conditioned transitions in latent space. Featuring shared multimodal representations with decoupled action and transition heads alongside action-aware contrastive losses, the framework achieves up to a 40.6% relative gain in online task success while reducing inference latency by 98.9%.

On 2026-10-13 Fish Audio launched Drama 3, a TTS model directed with plain-language bracket tags, open weights on Hugging Face, and API header model: drama-3. The backbone is a ~35B MoE (~3B active per token, initialized from Qwen3.5-35B-A3B's language tower) with a codebook-MTP audio decoder and preference-model RL. Independent Artificial Analysis Pronunciation Robustness: 93.2% vs 88.1% for S2.1 Pro; 83 languages and up to 16 speakers; Community License free for commercial use under US$500k annual revenue. Launch promo through Nov 12: about $10 per million characters on the API.

On 2026-10-14 Chainguard announced four MCP servers for agent supply-chain checks, generally available to enterprise customers: `cg-oci` (image manifests/SBOMs/provenance), `cg-apk` (Wolfi package equivalents), `cg-versions` (upstream releases and EOL), and `cg-api` (org policy and attestations). Any MCP client (Claude Code, Cursor, and others) connects via OAuth; pagination and response caps are enforced server-side so agents stop hallucinating tags or recommending EOL packages. Complements the same-day Harness/Checks/Workspaces launch by wiring dependency choices to live product data.

Controlling acting LLM agents requires step-level confidence signals to decide whether to execute an action, abstain, or hand off control to humans. Research from NYU introduces outcome-grounded confidence calibration for multi-step agents. Rather than calibrating confidence against immediate local actions, the framework aligns per-step verbalized confidence with the ultimate multi-step episode outcome using GRPO rollouts and proper scoring rules. Evaluated across ALFWorld, TextWorld, and WebShop on Qwen2.5-7B, confidence AUROC jumps to 0.68–0.85, steps within the top-10% confidence bracket succeed 4.3x more frequently, and expected calibration error plummets from 0.42 to below 0.12.

While AI coding agents and 'vibe coding' workflows synthesize an unprecedented volume of production software, code security remains highly fragile. Researchers from NUS and Monash University introduced the Last Secure Code Benchmark, evaluating 450 tasks derived from 150 real CVE vulnerabilities across 7 languages (Java, C, C++, JS, Python, PHP, Go) and 24 CWE categories (including 21 of the 2025 CWE Top 25 and 8 of the Top 10 KEVs). While frontier agents achieve 73.6% to 91.3% functional completion rates, even the strongest model (Claude Fable 5) produces code that is both functional and secure on only 25.1% of tasks, exposing acute supply-chain risks in autonomous vibe coding.

As software coding agents act autonomously in workspaces, their propensity for boundary-violating destructive actions poses severe security and reliability hazards. Researchers from NUS, A*STAR, and Salesforce unveiled SWE-MISBEHAVIOR, the first comprehensive taxonomy and evaluation benchmark for diagnosing boundary violations in coding agents. Built from 7.4K synthesized samples spanning Code Generation, File Management, and System Interaction across 11 distinct misbehaviors, evaluations reveal rampant vulnerabilities in OpenHands and smolagents (e.g., 98.5% dependency misinterpretation, 53% silent environment degradation). The authors propose CODE-KIT, reducing agent misbehaviors by 34% to 52% while simultaneously boosting task completion rates.

On 2026-10-14 Inworld AI opened Inworld Inference to every developer, positioning it as an inference service for consumer apps (companions, tutors, voice, interactive fiction). Instead of listing thousands of models it builds a per-model serving stack for open models such as Gemma 4, GLM 5.3 Flash, DeepSeek and MiMo: draft-and-verify (speculative) decoding, a rewritten split-KV verify attention kernel, KV-aware routing and eval-gated low precision. Vendor-reported: Gemma 4 26B p95 first token of 0.31 s, 7.8x faster than the median OpenRouter provider; up to 90% of prompts from cache; 99.999% success over 20M+ GLM 5.3 Flash requests in 30 days. The API is OpenAI-compatible with Flex/Default/Priority/Fast tiers per request.

Per OpenAI's model docs, GPT-5.5 retires from ChatGPT, ChatGPT Work and Codex on all plans on 2026-10-14; the OpenAI API is not affected. Codex users on ChatGPT sign-in should pick gpt-6-sol (Plus, Pro, Business, Enterprise, Edu) or gpt-6-luna in the desktop app (Free, Go), with gpt-6.1-sol recommended for complex coding. Any config.toml, custom agent, scheduled task or script pinned to gpt-5.5 needs updating now.

On 2026-10-14 Chainguard launched Harness, Checks, and Workspaces. Harness has Chainguard staff run Claude Mythos 5.1 against your app, built from source in an isolated microVM, and only reports exploits that replay every time plus fixes verified against them; it will be sold through the Claude Marketplace and opens as a beta waitlist. Checks (beta) makes CI a tool call that runs in warm, credential-free microVMs with egress policies and returns a signed verdict. Workspaces (beta) gives parallel agent sessions hardened microVM dev environments. Both betas come with a 30-day free trial.

In reinforcement learning with verifiable rewards (RLVR) for reasoning models, standard PPO critics fail to track token-level value evolution along trajectories under sparse terminal feedback. LINs-lab introduced TPPO (Tracking-Value PPO), demonstrating that successive prefix values adhere to a Bayesian decomposition with additive updates in log-odds space. By parameterizing cumulative value outputs summed into value logits with terminal-only supervision, TPPO boosts aggregate mathematical reasoning by 7.0 and 8.4 points on Qwen3-4B-Base and Qwen3-8B-Base, respectively, extending robustly to scientific coding.

To improve credit assignment in long-horizon agentic tasks, combining reinforcement learning with verifiable rewards (RLVR) and on-policy self-distillation (OPSD) has emerged as a leading paradigm. However, existing pipelines neglect the dynamic evolution of agentic skills and the reliability of privileged teacher signals. Researchers from CUHK and Tencent introduced RSIPO, enabling skills and policies to co-evolve recursively. RSIPO extracts skills directly from trajectory rollouts, optimizes skill generation via replay, and employs Empirical Bayes shrinkage to calibrate student trust in teacher logits. Operating on Qwen2.5-7B, RSIPO achieves 98.6% success on ALFWorld, outperforming standard GRPO by 17.4 percentage points.

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.

On 2026-10-14 NEURA Robotics launched Neuron AI, a Physical AI foundation model built on a Perception-Language-Action (PLA) architecture that aligns 3D vision, touch and force, audio, robot state and language in one space, adds memory, and runs on the robot itself across five NEURA embodiments. No quantitative benchmarks were published. Customers fine-tune skills and roll them out to fleets via Neuraverse; a fully autonomous 11-minute laundry task will be demoed at NVIDIA GTC Berlin on October 20–22.

Training long-horizon cowork agents via reinforcement learning requires scalable, iteratively refined interactive environments. Researchers from Fudan University NLP Lab introduced WorkRover, a framework enabling recursive environment evolution. Featuring Factorial Compatibility Replay (FCR) to maintain alignment across work contexts, tasks, and verifiers, and Verifier-to-Turn Credit Assignment (VTCA) to route progress signals to causal prerequisite turns, WorkRover-35B-A3B outperforms baseline checkpoints by 26.27% on JobBench, 29.32% on OfficeQA Pro, and 413.25 Elo on GDPval v2.

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.