Latest news, beside the leaderboard

CrowdStrike Intelligence found Claude Code session histories, ARTEX configs and a CLAUDE.md in attacker open directories, showing a likely Chinese-speaking, financially motivated actor used the China-built open-source agentic pentest framework ARTEX (DeepSeek v4.1-flash primary, plus GLM-5.3 and Grok 4.6) to breach loan-inquiry and employee mobile systems at South Korean financial institutions from late September to early October 2026.

Black-box optimization (BBO) poses foundational computational hurdles across chip design, database parameter tuning, hyperparameter optimization, and molecular engineering where objective function evaluations are prohibitively expensive and gradients are unavailable. Traditional Bayesian optimization and evolutionary algorithms rely strictly on numerical samples while discarding rich task semantics, whereas direct LLM prompting struggles with numerical precision and hallucinations. Researchers from Nanjing University's National Key Laboratory for Novel Software Technology (LAMDA Group) introduce AgenticBBO-Bench (arXiv:2610.12183), the first unified cross-domain benchmark for evaluating LLM agents in black-box optimization. Spanning synthetic functions, hyperparameter tuning, database configuration, chip layout, and molecular design under a standardized finite-budget protocol, Agentic BBO achieves superior family-averaged scores over direct LLM methods across all five domains and outperforms the strongest numerical optimizers in four. Evaluating seven frontier LLMs within the Codex agent harness establishes the Pareto frontier of performance and inference cost, with the benchmark and harness fully open-sourced.

Frontier language models generate incorrect conclusions with fluent, authoritative explanations, creating severe safety risks in autonomous systems where knowing when to defer to human review is essential. Traditional uncertainty quantification (UQ) methods rely either on compute-heavy repeated rollouts (e.g., semantic entropy across 10 to 20 samples) or supervised probes prone to confounding output length with true epistemic uncertainty. Researchers from TU Darmstadt and collaborating institutions introduce U-Space (arXiv:2610.09087), a mechanistic interpretability framework that uncovers evolving internal uncertainty within low-dimensional representations. By identifying semantic anchors of doubt and certainty in the unembedding matrix, U-Space constructs an orthogonal basis within the residual stream. The accompanying U-Lens projects intermediate token activations onto this basis, providing an interpretable, token-level uncertainty trajectory alongside an aggregated confidence score. Requiring zero training, zero labels, and zero repeated generations, U-Space outperforms established baselines under standard and length-controlled settings, with code fully open-sourced.

The emergence of reasoning models such as OpenAI o1/o3 and DeepSeek-R1 sparked widespread efforts to distill dense reasoning traces into smaller 7B and 14B student models, unlocking striking performance leaps on math and coding benchmarks. However, a fundamental mechanistic question has lingered: does On-Policy Distillation (OPD) inject new parametric factual knowledge into the student, or does it merely impart compositional reasoning skills? Researchers from Hong Kong University of Science and Technology (HKUST) address this mystery in a landmark study (arXiv:2610.09639). Utilizing a controlled synthetic framework across four models from three distinct architectures, the authors independently isolate factual memory from multi-step compositional ability. The experiments establish that reverse-KL OPD reliably transfers compositional problem-solving skills across unseen task structures, but transfers minimal factual knowledge. Replacing reverse-KL with forward-KL restores factual memorization, while student rollouts specifically refine multi-step logical execution. These findings demonstrate that on-policy distillation does not expand a model's knowledge boundary, but rather teaches it to mobilize and organize facts it already possesses.

Google's Gemini API release notes (Oct 8 entry) deprecate gemini-3.7-flash, with all requests now auto-routed to gemini-3.8-flash; gemini-3.5-flash is likewise routed to gemini-3.6-flash. The deep-research-pro-preview-12-2025 agent shuts down on Oct 23 and must migrate to the 04-2026 versions. Developers began receiving email notices on Oct 10.

On Oct 9 Ryan Dahl announced the whole Deno team is joining Cloudflare. Deno runtime gets one more year of monthly bugfix/security releases, then official development ends (it stays open source); Deno Deploy runs six more months before shutting down, with migration help to Workers; JSR continues. Focus shifts to the Workers/Durable Objects model and merging self-hosted celld into workerd, pitched explicitly at agent harnesses.

TypeSafe AI announced an $870M Series A at a $7.5B valuation on Oct 9, led by a16z with Sequoia and existing investor DCVC; Martin Casado joins the board. Its 'System One' decision model Jev launched Sept 15 and returns calibrated, typed decisions instead of text. TypeSafe says a third of the Fortune 500 use it; a16z's post says 25% — the figures differ.

Indie developer Nandakishor M open-sourced Vega (pip package vegaml, Apache 2.0): a frozen Qwen3.5-0.8B (4B also available) only reads, while a 57 MB trained engine models each allowed answer as a valley and rolls a damped ball into one, returning calibrated probabilities, conformal answer sets and an abstain flag. 73,728-token context and image input. The author's own paired tests beat hosted Jev 1.13.0 on tasks like phishing screening, while admitting it trails on reranking and multi-step reasoning.

On Oct 9 Cognition said Devin can now link any personal ChatGPT plan (Go, Plus or Pro). The GPT share of each session is billed to the ChatGPT subscription instead of Devin quota or on-demand credits, across Devin Cloud, Desktop and CLI. It requires a self-serve Devin Pro, Max or Teams plan (not Free or Enterprise), and falls back to Devin quota when the ChatGPT allowance runs out.

Modern AI agent systems and autonomous frameworks rely heavily on greedy search or reinforcement learning for self-evolution (such as harness synthesis, workflow routing, and prompt optimization). However, in non-convex and deceptive optimization landscapes, greedy pruning prematurely discards mutations that temporarily underperform, stranding systems in local optima. Researchers from Ant Group Research, Peking University, and Beijing Academy of Artificial Intelligence (BAAI) unveil Mara Chain (arXiv:2609.35855), an evolutionary optimization framework that rethinks failures as essential stepping stones. Mara Chain operates via three interconnected mechanisms: a stepping-stone archive combining semantic embeddings and execution traces for novelty preservation, a hierarchical mutator driven by backward failure attribution, and a Pareto-optimal non-dominated sorting selection module. On the AppWorld benchmark, Mara Chain achieves up to a 20.5% performance improvement over leading baselines (GEPA, ACE, SkillOpt-Lite) while requiring 65.5% fewer rollouts. On TerminalBench 2.1, it outstrips AHE and Meta-Harness by 20.2 and 22.5 percentage points respectively. The team has open-sourced the implementation under the AntOmniEvo repository.

As autonomous agents execute complex multi-step workflows across terminal consoles and web browsers, context window expansion and escalating token costs impose severe operational bottlenecks. Conventional compaction methods rely on text summarization or token pruning, both of which discard crucial state variables and induce semantic drift over extended sessions. Researchers from HKUST and Alibaba present REMORY (arXiv:2610.11287), introducing a residual memory paradigm designed for sequence-level context compaction. REMORY combines lossy natural language summaries with bounded sets of learned soft memory tokens, effectively forming a residual connection in the sequence dimension. While text summaries convey high-level narrative continuity, the soft memory tokens reconstruct lost fine-grained execution artifacts, API arguments, and intermediate environment states. On challenging benchmarks including BrowseComp and Terminal-Bench 2.1, REMORY retains 97% of full-context task performance while occupying only 5.2% of the original input token footprint, slashing redundant tool retries and loops by more than 40%.

Autonomous deep research agents (such as OpenAI Deep Research and Perplexity) represent a powerful leap toward automated knowledge synthesis. However, on open-ended and comprehensive inquiries, existing agents rely primarily on unconstrained search-and-generate loops. This leads to three systemic failures: redundant query loops, conflicting claims across sections, and runaway token expansion. Researchers from Tsinghua University and Renmin University of China unveil IOEDR (arXiv:2610.11566), a framework for Incremental Open-Ended Deep Research with a Structured Harness. IOEDR models the evolving research report as a dynamic hierarchical state graph coupled with a persistent, deduplicated evidence pool. Rather than executing open-ended browsing, the agent uses structured harness constraints to incrementally plan sections, identify specific evidential gaps, and trigger precision queries with dynamic termination conditions. On DeepResearch Bench, IOEDR achieves superior report coherence and verifiable fact density while reducing total token consumption by 33.2% and slashing search API requests by 61.4%.

On Oct 9 Prime Intellect shipped a Rust rewrite of Prime Agent, its open-source coding agent harness. A root orchestrator ran 2,209 sub-agents across 10,000+ Prime Sandboxes and ~228.7B GLM-5.3 tokens over two weeks, with differential TUI, harness and protocol parity tests gating every merge. Vendor-measured cold time-to-type dropped from 737.8ms to 55.8ms and post-startup memory from 607MB to 106MB; Windows (beta) and Homebrew installs were added.

Reinforcement learning (RL) has emerged as the defining post-training paradigm driving frontier foundation models toward autonomous self-improvement, yet scaling RL across multi-step agent environments and million-token sequences introduces profound stability and infrastructure hurdles. Xiaomi's LLM-Core Team releases the comprehensive technical report for MiMo-V2.6 (arXiv:2610.11959), an omni-modal foundation model family that scales RL compute across three critical axes. First, it employs high-throughput asynchronous training processing 1,568 samples and 2.7 to 3.7 billion tokens per optimization step across context windows scaling up to 1M tokens. Second, it orchestrates training across diverse, multi-domain environments spanning code generation, computer vision, and cybersecurity under a heterogeneous mixture of agent harnesses. Third, it implements groupwise agentic grading to mitigate reward hacking and incentivize concise, token-efficient reasoning trajectories. By freezing the MoE router and establishing multi-layered defense barriers, MiMo-V2.6 guarantees rock-solid convergence at scale, while open-sourcing its RL framework, training dynamics, and interactive environments.

Autonomous coding agents are increasingly leveraged to synthesize test suites for enterprise software. However, the standard practice of evaluating generated tests against a single reference solution overlooks alternative valid implementations and inflates perceived test quality. Researchers from Nanjing University and Hong Kong Polytechnic University unveil TestPrism (arXiv:2610.12289), exposing this systemic evaluation illusion. TestPrism comprises 300 test generation tasks sourced from 17 repositories and pairs them with 3,000 candidate implementations evenly divided into valid and invalid solutions. Its primary metric, the Joint Success Function, mandates that a test suite must simultaneously fail against buggy baselines, pass across all diverse valid implementations, and reject every invalid candidate. Evaluating fourteen frontier coding agent setups reveals that while models register a 59.67% success rate under legacy single-reference scoring, their actual Joint Success Function plunges to just 28.00%. To resolve these foundational flaws, the authors introduce TestHelix, an agentic synthesis harness leveraging peer cross-validation and recursive self-improvement to elevate Joint Success scores by up to 9.00 percentage points.

Long-horizon LLM agents accumulate expanding interaction histories that saturate GPU KV-cache memory and choke self-attention throughput. While sparse attention algorithms mitigate these bottlenecks mathematically, existing production serving engines (e.g. vLLM) are tightly coupled to dense tensor layouts, preventing heterogeneous sparse mechanisms from sharing infrastructure. Researchers from Harbin Institute of Technology (HIT) release SparseEngine (arXiv:2609.39068), an open-source, sparse-first inference engine built from the ground up for long-horizon agent workloads. SparseEngine establishes a shared lifecycle contract enabling 15 diverse sparse attention methods across four distinct families to control their custom KV representations while unifying memory management. It introduces Chain Cache to resume stateful KV-evicted contexts across multi-turn requests and Controllable Prefix-Cache Pruning to discard stale history without disrupting prefix matching. Empirical evaluations on long-horizon agent benchmarks show that SparseEngine delivers over 10x higher throughput under KV eviction, over 2.5x faster decoding throughput than vLLM at matched concurrency, and more than 2x end-to-end task speedups with zero degradation in model reasoning accuracy.

On Oct 9 Microsoft released Microsoft-Decision-1, a decision-scoring model that returns calibrated probabilities over fixed options instead of generating text, aimed at routing, classification, verification, agent guardrails and AI judging. Microsoft reports top accuracy across 36 blind benchmarks, ~35x lower P50 latency than GPT-6 Sol and a 1.3% decision-flip rate under perturbation. Available on Microsoft Foundry and OpenRouter, 32K context, $0.042/M input, free output.

On Oct 9 ClaudeDevs said every Pro and Max user on the Claude Code Projects waitlist has been let in. A project is a long-running coordinator conversation that spins each task into its own thread, usually a parallel cloud session on its own branch that opens a PR when needed; threads can also run locally via Remote Control. Still public beta; not yet on Team or Enterprise.

Long-horizon coding agents tackling complex software repositories depend on timely guidance, yet conventional LLM critics often degrade agent outcomes by misinterpreting in-progress exploratory commands or offering ephemeral advice that is never tracked. Researchers from Rutgers and Lehigh University unveil Opera (arXiv:2609.33987), an open-source verbal critic framework that treats diagnostic corrections as persistent notes tracked until root problems are definitively resolved. Opera orchestrates reviews through hybrid periodic and event-driven triggers, verifies proposed critiques against visible terminal artifacts, and monitors downstream trajectories to distinguish mere superficial compliance from genuine resolution. As a test-time intervention, Opera elevates baseline agent resolve rates by 12.4 percentage points on Terminal-Bench 2.1, 15.0 on SWE-Bench Pro, and 8.9 on DeepSWE v1.1 across four foundation backbones. Furthermore, fine-tuning Qwen3.5-9B on Opera-guided rollouts yields a 10.2 percentage point gain on held-out repositories without any test-time critic, preserving stability across heterogeneous execution harnesses (such as OpenHands to Terminus-2).

Endowing artificial agents with general intelligence requires inferring environment dynamics and refining world models continuously in the wild. Fine-tuning models directly is prohibitively expensive and precipitates catastrophic forgetting. Researchers from University College London (UCL) and Huawei Noah's Ark Lab UK present Memento 3 (arXiv:2610.11794), a framework enabling frozen LLM agents to achieve model-based Recursive Self-Improvement (RSI) purely via external memory. The agent maintains a natural-language rulebook representing its evolving hypotheses of environment physics, which is compiled into executable code for fast simulation and action planning. Governed by a continuous observation-reflection-revision-compilation-verification loop, the system leverages prediction errors to refine rules. Edits are merged only after passing LLM semantic alignment audits and deterministic cell-exact replay verification. On the rigorous ARC-AGI-3 abstraction benchmark, Memento 3 solves 100% of levels across all 25 public environments, scoring a perfect 100.0 Relative Human Action Efficiency (RHAE) while utilizing only 44% of human action counts. In Atari Pong, a synthesized feedback controller shuts out opponents 21:0 across three evaluation seeds with zero test-time LLM inference.