Latest news, beside the leaderboard

On 2026-10-12 LightOn AI released ColGrep-Agent-2B (fine-tuned from MiniCPM5-2B, Apache-2.0): a local 2B code-localization agent that searches with ColGrep multi-vector retrieval plus a read-only terminal, then returns file paths and line ranges to a frontier coding agent. On 52 code-search questions across 13 unseen repos (624 Claude Code / Sonnet 5.5 sessions), delegating search cut API tokens ~34.3% and cache-aware cost ~40% vs grep, with 80.8% vs 83.3% accuracy (not statistically significant). On SWE-bench Lite localization it posted 19.92 line F1—best among models evaluated in the same ColGrep harness, including several >10× larger. Ships brew/cargo/GGUF installers and hooks for Claude Code, Codex, and OpenCode.

On 2026-10-12 SpaceXSI launched The X Subscription: Free $0, Plus $8, Premium $30 (Grok Bot, 4× usage), Super $100 (15×), Ultra $200 (50×), and Teams from $30/seat. Grok and Cursor share a weekly-refreshing usage pool; third-party models in Cursor bill at API list price. Existing subscribers can keep or switch plans; rollout is gradual.

On 2026-10-12 Microsoft EVP Scott Guthrie announced that Azure Maia 200 is in production for Copilot-class workloads, and third-generation Maia 300 targets long-context reasoning, latency-sensitive chat, and multi-step agentic inference. A 96-accelerator / 167 kW rack is claimed to deliver up to 7× Maia 200 rack-scale token throughput and >2.5× performance per watt, plus ~40% more tokens per megawatt vs leading GPU systems. Each accelerator has eight HBM4 stacks (~288 GB / 18 TB/s), TSMC N2P compute chiplets, and Ethernet-based AI transport scaling to a 12,288-accelerator superpod. Production deployments begin in 2027; more at Ignite 2026.

On 2026-10-12 Anthropic published a research post claiming substantially stronger defenses against illicit distillation of Claude: early evidence of ~90–99% less data harvested by distillers, and ~10–100× fewer reasoning tokens extractable on-platform after new safeguards. Methods include preserved thinking (harder context rewriting for newer API accounts), org-scoped reasoning transcripts, a two-stage classifier against chain-of-thought extraction (<0.1% claimed false-positive rate), bans/limits, and resumed charging for blocked low-FPR requests. The post cites cyber-capability distillation into models such as GLM-5.3 and urges other frontier labs and platforms to share the load.

On 2026-10-12 AMD introduced Helion, an OCP-based open rack blueprint for agentic AI infrastructure building on Helios. Hyperscale CPU-only racks with 6th Gen EPYC can reach up to 32,768 cores for agent execution tied to separate inference; enterprise all-in-one racks combine EPYC, Instinct MI350P and general-purpose compute. Open hardware/software plus an integrated token router for frontier and open-source model calls; AMD plans to contribute the spec to OCP, with partner configurations targeted for November 2026.

On 2026-10-12 Anthropic (William Zhou, Daniel Freeman) published [research](https://www.anthropic.com/research/self-replicating-ai) arguing small models may be months from autonomous weight-based replication on commercial IoT. Models ≤30B already complete many replication steps; in simplified evals **GLM-5.2** finishes end-to-end. A demo ran **Gemma4-26B-A4B** across 15 ~$55 Android TV boxes via pipeline parallel (~0.39 tok/s). Extrapolating Epoch Capabilities Index gains, small models may hit the replication floor by end of 2026.

Continual learning without weight fine-tuning represents a frontier in LLM agent research. When an agent acts purely by writing and executing Python scripts inside an execution sandbox, every thought, hypothesis, and plan leaves a verifiable trace on disk. In a newly released study (arXiv:2610.11450), researchers subject coding agents to ARC-AGI-3—a suite of instruction-free abstract visual reasoning environments where each sequential level introduces novel causal dynamics. Developing a fine-grained measurement protocol, the authors trace thoughts across file artifacts across seven full evaluation runs spanning three foundation backbones from two major model families. The empirical findings reveal that scripts written for one level are almost never invoked in subsequent levels (only 33 of 630 cross-task references survive), as agents deliberately abandon 74% of scripts at task boundaries to purge level-specific overfitted state while rewriting core rules. Notes are maintained in an append-only fashion, with conflicting claims arbitrated against raw execution logs to achieve selective rather than catastrophic forgetting. Crucially, the study identifies carrying hard-coded variables across task boundaries as the single most devastating failure mode of autonomous agents.

Rendering dense, legible, and spatially accurate text within generated images is critical for commercial graphic design, advertising posters, and e-commerce visuals. However, existing benchmarks predominantly test isolated words or simplistic English phrases, failing to capture the complexity of rendering multi-region, bilingual text layouts. LINs-lab introduces UltraText Bench (arXiv:2610.09823, code: github.com/LINs-lab/UltraText_Bench), a comprehensive bilingual benchmark designed to stress-test prompt-only dense visual text rendering. The suite contains 432 human-reviewed prompts distributed equally between English and Chinese across 24 real-world categories and three difficulty tiers (L1 to L3), with each prompt prescribing exact text for 4 to 12 distinct visual regions. Utilizing the Q-Judger multimodal evaluation model, UltraText Bench assesses 24 state-of-the-art model configurations across text fidelity, text clarity, spatial layout, and scene aesthetics. Empirical results expose key architectural trade-offs: Turbo distilled variants gain 3.81 clarity points but sacrifice 14.76 fidelity points, while flagship models like Qwen-Image exhibit steep performance degradation under L3 dense workloads (dropping from 86.50 to 42.86), with code and evaluation suites fully open-sourced.

Autonomous research agents hold the promise of sustaining AI self-improvement by iteratively proposing hypotheses, tuning models, and refining post-training pipelines. While agents demonstrate competence on text-only code tasks, their capacity to execute multimodal post-training across images, audio, and video remains uncharted. Researchers introduce MMPostTrainBench (arXiv:2610.05398), the first benchmark designed to evaluate autonomous research agents conducting multimodal post-training across eight tasks spanning image, audio, video, joint audiovisual comprehension, and image-grounded software repair. Agents operate from a common base model under fixed budgets with development feedback before independent evaluation. The audit reveals sobering realities: across all tasks, 52.1% of agent-submitted models fall below the baseline, exhibiting severe non-target capability regressions, and agents fail to reliably select the best candidate (trailing top models by up to 5.38 percentage points). To resolve multimodal research drift, the authors introduce MMResearch, a framework integrating media-grounded evidence and hierarchical memory. Integrated into coding agent runtimes, MMResearch improves submitted model accuracy by up to 7.75 percentage points for Claude Code (Claude Opus 4.8) and 2.33 points for Codex (GPT-5.6).

As autonomous software engineering agents and command-line assistants evolve from single-turn tool invocation toward complex, multi-stage delivery, the scarcity of high-quality, long-horizon terminal interaction data has emerged as a fundamental bottleneck. Researchers from USTC, Alibaba Group (AMAP), and SUSTech introduce TerminalHorizon (arXiv:2610.13166), a fully automated data synthesis engine for training long-horizon terminal agents. Addressing standard pitfalls where synthetic environments oversimplify core challenges and eliminate the need for exploration, TerminalHorizon extracts workflows from real-world tasks, pinpoints omitted steps through independent execution loops, dynamically reconstructs environments based on observed evidence, and organically scales task complexity across multiple stages under unified objectives. Synthesizing 1.5K environments and 2,852 trajectories averaging 52.8 turns, the authors train TerminalHorizon-27B (fine-tuned on Qwen3.5-27B). The model achieves 55.8% on Terminal-Bench 2.0 (+14.2 points over base Qwen3.5-27B), sweeping 12 benchmarks across coding agents, production environments, and reasoning with project datasets open-sourced.

As LLM agents are deployed on long-horizon, end-to-end open web tasks such as deep research, multi-vendor comparison, and cross-site navigation, existing web access paradigms suffer severe fragmentation: search-augmented agents discard interactive controls, browser-based agents incur prohibitive context token consumption, and filesystem-based tools (Fetch-then-Explore) lose crucial interaction contexts when selectively grepping text. Researchers from Yonsei University introduce Web$hell (arXiv:2610.13416), a unified interface standardizing web access. Anchored by WebContract (establishing consistent schemas for addresses, values, and associations), WebCompiler (deterministically transforming HTML semantics into operable information and action objects without LLM inference overhead), and WebCache (maintaining persistent, stateful page revision trees), Web$hell unifies passive discovery with active interaction. Evaluated across six benchmarks spanning Web Search (WideSearch, SealQA), Web Navigation (WebArena-Verified, WebWalkerQA), and Deep Research (DeepResearchBench, DeepWeb-Bench), Web$hell achieves state-of-the-art performance across all tasks. Crucially, it slashes Deep Research turns from 52.4 and 30.9 down to 17.6 (-43% to -66%), maintaining flat cost curves even when subjected to 400 injected distractor pages.

Empowering AI agents to autonomously achieve Recursive Self-Improvement (RSI) through ongoing environmental interaction is a cornerstone of artificial general intelligence. However, practical RSI pipelines are severely constrained by two fundamental bottlenecks: scarce and expensive ground-truth feedback, and the degradation of predictive world models as evolving policies induce out-of-distribution drift. Researchers from HKUST, Shanghai AI Laboratory, Peking University, and CUHK introduce RSI-Anything (arXiv:2610.13912, code: github.com/Gen-Verse/RSI-Anything), a domain-general framework unlocking recursive self-improvement via predictive world models. Built upon the JEPA-Anything architecture, RSI-Anything establishes a disciplined six-stage paradigm: Propose → Predict → Select → Evaluate → Learn → Retain. By enforcing rigorous compatibility and safety gates governing policy-world model retention, the system continuously converts interaction logs into persistent capability upgrades. Across diverse domains, RSI-Anything drives substantial performance gains: boosting automated code repair success from a 26.3% baseline to 31.2% (+4.9 points), reducing GUI navigation action footprints by 39.4%, more than doubling laboratory planning scores from 21.80% to 46.01% over five iterations, and achieving a 96.15% chemical yield, with codebase fully open-sourced.

On 2026-10-12 Z.ai introduced GLM-Legal, built on GLM-5.3-Flash with legal mid-training and RL across tens of thousands of legal environments, co-trained with its agent harness. Vendor-reported: 19.2% Harvey LAB all-pass (base 6.7%); LegalBench and PLawBench claimed at parity without scores. No open weights; API by application; AutoClaw support coming.

On 2026-10-12 @OpenAIDevs highlighted ChatGPT desktop browser Annotations extensibility (`document.oai.annotation`): sites can customize selection targets, metadata and preview controls, and pair them with WebMCP site tools so agents receive the user's selected page context and can act through the site's own capabilities. Docs target DevDay 2026+ desktop; Enterprise/Edu need admin policy.

On 2026-10-12 OpenBMB released MiniCPM-V 4.7, an Apache-2.0 open multimodal family with a 35B-A3B variant (the series' first MoE) and a 1B variant. Its Lookahead Expert Scheduling (LES) prefetches experts across layers and, combined with layer-wise mixed-precision quantization and DSpark speculative sampling, lifts SSD-offloaded decoding on an iPad Pro M4 (16GB) from 2.97 to 17.39 tokens/s (5.86x), per OpenBMB. Weights are on Hugging Face, ModelScope and Ollama, with vLLM, SGLang, llama.cpp and iOS/Android/HarmonyOS support.

On 2026-10-12 Meta introduced Muse Wonder and Muse Wonder Full, its most capable models yet for personal and professional agentic work. Official posts highlight Active Memory (carry and retrieve what matters across conversations) and Active Focus (persist through hard tasks, recover from errors, clarify and confirm before consequential actions). Muse Wonder is available today via the Meta Model API and Muse Code; Muse Wonder Full is available today for Muse Code subscribers. Third-party Artificial Analysis Intelligence Index (max reasoning) lists Wonder Full at 56 and Wonder at 52 (Spark 1.3 Max was 48).

Tackling long-horizon problem solving and scientific discovery requires computation to compound across successive attempts. While humans build upon partial proofs, failed experiments, and prior hypotheses, current test-time compute paradigms (e.g., MCTS and Best-of-N) isolate search trajectories into independent silos, discarding intermediate reasoning artifacts upon branch failure. Researchers introduce compounding inference and its open-source instantiation GitSwarm (arXiv:2610.04862). GitSwarm establishes an asynchronous multi-agent coordination architecture where homogeneous LLM agents collaborate through structured persistent memory hosted in a shared, branchable Git repository. Intermediate artifacts are preserved as atomic commits, while explicit semantic dependencies link cross-branch contributions, enabling agents to inspect, branch, verify, and synthesize past discoveries. Evaluated on rigorous long-horizon benchmarks, GitSwarm paired with GPT-5.5 solves all 30 problems on IMOProofBench-Advanced in a single execution run, and scores 79.4% on ProgramBench compared to 65.1% for the strongest reported baseline under identical budgets.

As autonomous coding agents undertake multi-hour, long-horizon software engineering tasks, humans transition from granular copilots to high-level overseers. However, auditing agents across thousands of shell executions and fragmented code edits creates severe cognitive overload: overseers struggle to locate decisive actions and identify consequential decisions requiring human verification. Researchers introduce AgentMonBench and the Evidence-Grounded Behavior Graph (EBG) framework (arXiv:2610.06406, code: github.com/zhk-lab/EBG, site: zhk-lab.github.io/agentmonbench-ebg). AgentMonBench assesses oversight monitors across requirement alignment and decision risk awareness. EBG is a training-free methodology that structures source-linked execution evidence into semantic behaviors, assembling their dependencies into a directed causal behavior graph with task-oriented views. Tested across eight foundation models, EBG consistently outperforms raw-context baselines in identifying consequential decisions and localizing supporting evidence across varied trajectory scales, with code and datasets fully open-sourced.

Reconstructing scene-aligned, high-fidelity 3D assets from casual photos is foundational for robotics and spatial computing. However, casual captures suffer from sparse viewpoints, optical blur, and severe occlusions, causing standard 3D diffusion models to produce warped unobserved backsides and incorrect 6-DoF poses. NVIDIA Research introduces GATOR (arXiv:2610.11215, project: research.nvidia.com/labs/lpr/gator), a generative and agentic framework that reconstructs complete textured 3D objects and their scene-relative poses from one or more casual images. GATOR couples patch-aligned RGB, target masks, and pointmap features through a local modality mixer, synthesizing initial textured geometry via stage-specific adapters. Crucially, the generated asset initializes a multimodal agent that iteratively inspects, refines, and aligns the asset through an observation-guided edit-render-review loop. Across synthetic benchmarks, cluttered tabletops, and real indoor scenes, GATOR achieves state-of-the-art geometric fidelity and instant simulation readiness with superior inference efficiency.

Anomaly published the OpenCode 2.0 launch note on Oct 12: a single auto-started local service shared by the TUI, web and desktop apps; a rewritten server API with the new @opencode/client; @opencode/codemode, which turns MCP servers and OpenAPI endpoints into functions the model composes in scripts (all MCP servers opt into Code Mode by default); background subagents and background shell commands; 80+ internal features rebuilt as hot-reloadable plugins; and @opencode/sdk for embedding the same engine as a library. V1 plugins and the old server API are not compatible.