arXiv
AI overview
Sign in and the AI will write an overview from our coverage.
Headlines · 24
- Claude Mythos 5 designs more efficient quantum circuits
Anthropic said its STEM Fellows Program had Claude Mythos 5 design quantum circuits from plain-language specifications alone, working largely on its own across eight circuits drawn from chemistry, optimization, physics simulation and number theory, with an automated evaluator scoring each design. According to Anthropic, the resulting circuits improve on the best published ones, in some cases by an order of magnitude or more. The work was led by Andi Gu with mentorship from Anthropic quantum physicist Alek Kemeny, in collaboration with Susanne Yelin and Mikhail Lukin of Harvard University (Lukin also at QuEra Computing), and the full paper is on arXiv. Anthropic argues that more efficient circuits reduce the qubit and runtime resources quantum algorithms need, which could bring practical applications in chemistry and materials science closer.
Anthropic (X) · 🔥 22 - Tencent open-sources DSpine, a parallel drafter for speculative decoding
Tencent published inference code and open weights for DSpine as tencent/Qwen3-8B-DSpine-b16 on Hugging Face under Apache 2.0. DSpine is a parallel drafter for speculative decoding: at every draft layer each position receives the predicted feature of its predecessor, so the causal chain unfolds over network depth while all positions are still drafted in parallel. The model card says the paper was released on arXiv in September 2026 and the inference code in October 2026, with training code and a Qwen3.8-27B version still on the roadmap. In the team's comparison on Qwen3-8B against DSpark, DFlash2, DFlash and Domino (all retrained with the same configuration on Open-PerfectBlend), DSpine reports the best average acceptance length, 7.83 at T=0 and 6.98 at T=1, though these figures are self-reported by the releasing team.
腾讯混元 · Hugging Face · 🔥 8 - Microsoft study: agent long-term memory may hurt; Fork-and-Flush resets context
A Microsoft Research paper posted to arXiv on October 5, “Fork-and-Flush: Escaping Idea Basins in Autoresearch Agents,” reports that autoresearch agents can get stuck in “idea basins”: running the same agent on the same task eight times yielded final scores from 68.3 to 100.0 (median 81.9), and throwing extra compute at the weak runs did not close the gap. The proposed Fork-and-Flush method copies the agent into parallel branches that inherit the full workspace — code, files and experiment results — but start with a cleared chat history, then keeps the best-scoring branch after a fixed period. Across 13 long-horizon tasks under the same compute budget, the normalized average score rose to 0.78 versus 0.47 for a single run and 0.54 for best-of-four, with standard deviation falling from 0.26 to 0.16; in ablations, flushing alone gave roughly a 22% gain while forking and flushing together worked best. Tests with Qwen3.5 at 4B, 9B and 27B on MAX-3-SAT showed run-to-run variance peaks near the model’s capability edge, at about 0.5 attainment, suggesting the effect may persist as models improve.
36氪 人工智能 · 🔥 6 - MRTPO puts each reward on its relevant tokens in multi-reward RL
A preprint submitted to arXiv (arXiv:2610.21439, 15 Oct 2026) proposes MRTPO, a method for credit assignment in multi-reward reinforcement learning. The authors note that reward-decoupled normalization (GDPO) normalizes each reward separately but still spreads every reward's advantage evenly over the whole response, even though each verifier checks only a few positions. MRTPO instead replays the policy's own response with and without a given reward's verifier information and uses the per-token Jensen-Shannon distance between the two next-token distributions to measure each token's relevance, reassigning that reward's advantage accordingly, without auxiliary models or token-level annotations. The authors report faster convergence and better results than the compared methods on text-to-SQL, multi-constraint instruction following and tool calling; the paper is 15 pages with 8 figures.
Hugging Face · Papers · 🔥 0 - FDE-Gym benchmark: agents score 0.56 on real-world delivery tasks
A new paper listed on Hugging Face's Papers feed introduces FDE-Gym, a benchmark of 95 end-to-end delivery tasks drawn from 20 customer scenarios across 18 sectors and adapted by industry experts from real delivery projects. In simulated customer environments, agents interact with scripted stakeholders and work on production-like systems, scored by how much they improve the customer's own success metric (0 for an unchanged system, 1 for a reference solution), with penalties for messages, trial runs and waiting time. The authors report that in the realistic setting, where unstated constraints must be discovered, the best configuration reaches a mean FDE-Gym score of 0.56 and resolves only 14 of the 95 tasks; even on the 23 tasks where it obtained every key fact from stakeholders, only 10 reached a normalized business uplift of 0.8 or above.
Hugging Face · Papers · 🔥 0 - SpecAgent has weak models draft tool-call trajectories and strong models verify
An arXiv paper introduces SpecAgent, a collaboration protocol that turns a strong model from the author of each agent step into the reviewer of steps drafted by a weaker model. The weak model executes read-only tool calls while drafting so each step is grounded in real observations, and stops at the first action that would modify the environment; the strong model then verifies the drafted segment in a single prefill pass, commits the longest prefix it judges semantically valid, and repairs the first rejected step. The authors report SpecAgent comes within one point of the strong model's success rate on τ²-Bench while roughly halving memory traffic and wall time, recovers 44–75% of the quality gap on AutomationBench and VitaBench, and in an edge-cloud MCP-Atlas deployment closes 53–75% of the weak-strong quality gap with the strong model generating at most half as many tokens.
Hugging Face · Papers · 🔥 0 - Four Gemini 4 Argon agents beat Minecraft in 27 hours
Google DeepMind highlighted an arXiv paper (arXiv:2610.22319, submitted Oct 15, by Maxence Faldor, Satinder Singh and Roberta Raileanu) that placed four Gemini 4 Argon agents in a sandboxed Minecraft survival world with one goal: defeat the Ender Dragon. According to the paper, each agent controlled a player through the open-source Mineflayer library and coordinated via in-game chat and shared memory, planning, writing and debugging every program themselves; the team wrote 2,783 programs and killed the dragon in 27 hours and 14 minutes, dying 605 times, while researchers watched on a live dashboard without intervening and no roles or game logic were assigned by humans. The authors say this is the first end-to-end autonomous completion of Minecraft by AI agents using only the standard Mineflayer library, without pre-scouted seeds or human-written game-specific tools, in a randomly generated Easy-difficulty world, and argue the same long-horizon coding and multi-agent coordination is needed for real-world software engineering and enterprise tasks.
Google DeepMind (X) · 🔥 8 - last30days agent skill adds native Codex plugin, new sources
The open-source AI agent skill last30days has updated to v3.11.1. According to the project, it has merged 175 pull requests since the v3.3 announcement in May — 122 of them from 52 community contributors across 15 releases. The skill lets an agent search Reddit, X, YouTube, Hacker News, Polymarket, GitHub and other platforms in parallel, rank results by upvotes, likes and money at stake, and synthesize one brief. This update adds a first-class OpenAI Codex plugin with renderer-aware citations, free keyless sources including arXiv, Techmeme and Digg, a rebuilt X search pipeline, and a doctor command that checks each source's health.
GitHub Trending · Python · 🔥 0 - Survey proposes two-axis framework for multimodal agent memory
A survey submitted to arXiv on 14 Oct 2026 (arXiv:2610.19366) by Minghao Guo and 79 co-authors reviews how multimodal agents handle memory. It organizes the field along two axes: Axis I covers how multimodal information is formed and represented (for example modality-specific embeddings, textual and structured records, latent and parametric representations), while Axis II splits memory management into T1 static memory, T2 dynamic memory and T3 self-evolving memory, describing how management policies adapt as experience accumulates. From that framework the authors synthesize recurring design patterns and principles, including balancing source fidelity against abstraction, preserving temporal validity and provenance, grounding memory use in recoverable evidence, and evolving memory policies over time, and they outline open challenges and future directions.
Hugging Face · Papers · 🔥 0 - Temporal Flow Policy adds physical time to robot policy memory
An arXiv robotics paper submitted on 14 Oct 2026 introduces Temporal Flow Policy (TFP), treating physical time as a missing dimension of robot policy memory. The authors argue that existing policy memories evolve with observation sequences, implicitly treating every update as the same amount of physical time, even though identical observation histories can span different durations and require different actions. TFP learns memory dynamics that retain past events and advance the memory state over measured time intervals, conditioning diffusion and flow-matching policies through that state and training with their native imitation objectives. In controlled comparisons with Diffusion Policy and π0.5, the authors report that TFP achieves higher average success than recurrent and state-space memories given the same intervals, and than a variant that feeds the interval to a learned gate; they also report gains on real-robot timing/memory tasks and LIBERO evaluations.
Hugging Face · Papers · 🔥 0 - SwarmRSI evolves scientific agent swarms for recursive self-improvement
An arXiv paper submitted on 14 Oct 2026 introduces SwarmRSI, a framework that evolves scientific agent swarms so agents improve their research capabilities through their own experimental experience. It combines Feedback-Guided Swarm Mutation, which refines hypotheses and their implementations using feedback from other agents, with Complementary Swarm Recombination, which reuses retained hypotheses and complementary program components so partial results can support new research; Recursive Swarm Evolution ties both into a feedback loop. The authors report state-of-the-art results on OpenProblems, NatureBench and MLE-Bench Lite, say accumulated research experience improves performance on held-out tasks, and note that biologists reproduced its programs for follow-up biological analyses.
Hugging Face · Papers · 🔥 0 - Paper claims AI models can provably hide arbitrary capabilities
An arXiv paper titled "AI Models Can Provably Hide Arbitrary Capabilities" appeared on Hacker News' AI section, drawing more than 100 points and an active comment thread. Judging from the title alone, the authors argue that models can hide arbitrary capabilities in a provable way; the source provides no abstract, authors, affiliations or experimental details, so the claim remains unverified and should be attributed to the paper's authors.
Hacker News · AI（100+ 分） · 🔥 0 - DeltaTTG trains multi-turn agents with a Token-to-Go progress meter
An arXiv paper introduces DeltaTTG, which learns a reusable progress signal through a shared progress meter for reinforcement learning of multi-turn agents; its supervision target, Token-to-Go (TTG), is the number of tokens remaining to be generated until task completion, used as an observable proxy for remaining completion effort. The method uses low-cost supervision from recorded trajectories of diverse policies without task-specific progress rules, and during actor training, changes in the frozen meter's estimates redistribute outcome-derived advantages across steps without altering the reward. On ALFWorld, PaperSearchQA and BFCLv3, DeltaTTG improves the average score over GRPO by 5.7 and 7.2 points with Qwen3-4B and Llama-3.1-8B-Instruct respectively, outperforms all compared RL baselines on average while keeping actor-training GPU-hours close to GRPO, and the authors also report improved sample efficiency and out-of-domain performance; code is said to be available.
Hugging Face · Papers · 🔥 0 - WorldRecon: Generative multi-view mesh reconstruction for large-scale 3D scenes
A new arXiv paper introduces WorldRecon, a generative multi-view reconstruction framework that recovers complete, large-scale 3D scene meshes from unposed RGB images. It uses a coarse, noisy global point cloud and camera poses estimated from the inputs as a global scaffold, partitions the scene into overlapping spatial blocks, and equips the native 3D generative model TRELLIS.2 with reconstruction-based conditioning to produce sparse structure, detailed geometry and PBR materials from each block's local point-cloud crop and relevant image observations. To keep scenes coherent, the authors add neighbour-aware latent inpainting and a causal world-generation procedure that propagates structural, geometric and material context across blocks. The paper reports state-of-the-art results, reducing Chamfer distance by 14.9% and 43.1% and improving F-score by 17.3% and 12.0% over the strongest baselines on synthetic and real-world indoor scenes respectively, with strong claimed sim-to-real generalization.
Hugging Face · Papers · 🔥 0 - JevGPT: a dual-system architecture for autonomous research
A paper submitted to arXiv on 13 October 2026 introduces JevGPT, a dual-system architecture for autonomous research that couples fast evidence processing with deliberative research planning. System Two sets evidence objectives and integrates findings, while System One locates supporting passages, assesses relevance and guides further inspection, with independent source judgments run in parallel. The authors report that on fixed-corpus question answering and document ranking JevGPT achieved leading task performance with faster completion and lower cost than comparably effective research agents; these results come from the paper's own experiments and have not been independently verified.
Hugging Face · Papers · 🔥 0 - AReA: Training-Free Attention That Boosts High-Resolution Diffusion Transformers
Researchers posted AReA, a training-free adaptive receptive-field attention mechanism for high-resolution image generation with Diffusion Transformers (DiTs). The paper argues uniform full attention is the common cause of both the quadratic cost and the structural distortions and lost detail seen at higher resolutions, because attending to all tokens at every step pushes the model beyond the token counts and relative distances seen in training. AReA instead lets the effective receptive field evolve from coarse to fine along the denoising trajectory: early steps apply cross-scale coarse attention over downsampled keys and values to capture global context, while later steps apply local attention within progressively shrinking windows. The authors report that, with no extra parameters or training, AReA consistently beats state-of-the-art training-free baselines across target resolutions and models while also being faster and more efficient.
Hugging Face · Papers · 🔥 0 - arXiv caps submissions at two per month as AI paper flood strains moderators
arXiv began enforcing a submission rate limit on October 1, 2026, allowing each submitter at most two papers per calendar month while keeping the existing cap of three active submissions at any time; the rule covers all categories, counts only the person submitting, and rejected papers still consume a slot unless deleted before publication. arXiv calls it a temporary measure: September 2026 brought 40,363 submissions, up from 20,569 two years earlier and 9,869 in September 2016, with cs.AI submissions growing more than sixfold in two years and generating nearly 9,000 support tickets. Editorial Advisory Council chair Thomas G. Dietterich said a small share of authors submit many low-quality, thin or AI-written papers, consuming a disproportionate share of volunteer moderators' time and delaying others by days or weeks; arXiv left Cornell University on July 1, 2026 to become an independent nonprofit and received $17.2 million in pledges in late September, partly to handle AI-generated content.
The Decoder · 🔥 0 - Open-source AI paper tool checks citations against full texts
Individual developer Maoshu (GitHub user mmlong818) released Open Academic Paper Gen, an open-source multi-agent workflow that drafts papers in nine stages: scoping the research area, literature search, cleaning and filtering, trend analysis, finding an angle, outlining, writing, verification and export. It searches OpenAlex, Crossref, Semantic Scholar and arXiv, then expands the pool by following references and citing papers. Its distinguishing feature is three layers of citation checking: whether a citation exists in the collected pool, whether the paper's identity matches Crossref or doi.org records, and whether the full text actually supports the claim; when full text is available it locates the relevant passage and, if page data exists, flags the PDF page, falling back to abstracts or excerpts otherwise, and marking insufficient evidence as "undetermined" rather than passing it. The default mode pauses at major stages for human review, with a fully automated option also available; the project supports OpenAI and Zhipu GLM, though claim verification, uncited-statement checks and simulated review still require an OpenAI API key, and exports cover Markdown, LaTeX, RIS and BibTeX.
36氪 人工智能 · 🔥 0 - ByteDance Seed paper ties DeepSeek long-context swings to KV-cache phase sensitivity
Per DoNews, ByteDance's Seed team posted an arXiv paper in late September 2026 arguing that "phase sensitivity" introduced by chunked KV-cache compression is the main driver of long-context retrieval fluctuations in DeepSeek-V4 series models. Covering open-weight models including DeepSeek-V4-Flash, V4-Pro and V4.1-Flash, the work says compression windows impose token phase coordinates, so the same information can be retrieved with accuracy differing by as much as 40 percentage points across phases — a periodic degradation that averaged benchmark scores tend to hide. The team calls for better cache-compression designs to improve long-context stability.
DoNews · 🔥 0 - Set theorist blasts OpenAI's Partition Principle preprint as unreadable
Set theorist Asaf Karagila responded to OpenAI's preprint claiming the Partition Principle does not imply the Axiom of Choice, calling it muddled, oddly structured, loosely termed and poorly referenced — the sort of submission that should receive a desk rejection. He said OpenAI released hundreds of incomprehensibly written "solutions" without expert input, effectively shifting the burden of sorting and verification onto mathematicians. He therefore will not send Sam Altman the promised bottle of whisky and will not spend time reading the paper.
Hacker News · AI（100+ 分） · 🔥 0 - 'Foundations of Large Language Models' book posts v4 revision
The arXiv book "Foundations of Large Language Models" by Tong Xiao and Jingbo Zhu has been updated to version 4, described by the authors as a set of minor corrections. The book has six main chapters covering pre-training, generative models, prompting, alignment, inference and reasoning, and is aimed at college students, professionals and NLP practitioners as a reference on foundational concepts rather than the full range of cutting-edge techniques.
Hugging Face · Papers · 🔥 0 - Tetris3D generates physically coherent 3D scenes from a single image
An arXiv paper introduces Tetris3D, a generative framework for single-image 3D scene reconstruction that explicitly conditions each object's generation on the geometry and physical relationships of surrounding objects, keeping shape and pose geometrically and physically plausible within the scene. The authors report that Tetris3D recovers coherent object shapes and poses even when interacting regions are occluded, and achieves state-of-the-art results in generation quality and physical stability. The work also releases ComOb, a physics-simulation-based dataset of 1.2M scenes with per-object meshes and pairwise physical relation annotations.
Hugging Face · Papers · 🔥 0 - UniSkill lets one policy co-evolve with its own skillbank
An arXiv preprint introduces UniSkill, in which a shared policy interacts with the environment and proposes skillbank edits (Add, Update, or No Edit) drawn from the resulting trajectories, so the policy and skillbank co-evolve. Contrastive action feedback supplies an actor-alignment signal by measuring how swapping the retrieved skill for a proposed one changes the actor's action log-likelihood gap between previously collected successful and failed trajectories on the same task, avoiding a new rollout per proposal; skill-edit support regularization is added to preserve exploration. The authors report 98.4% success on ALFWorld and 84.7% on WebShop with stable joint training, and say the approach still works with a smaller backbone policy; code is released.
Hugging Face · Papers · 🔥 0 - SkillForge co-evolves agent skills and policy via a fitness-driven skill lifecycle
A NeurIPS 2026-accepted arXiv paper introduces SkillForge, an agentic RL method that evolves a skill library through a fitness-driven lifecycle of trial, active, stable and retired states, so skills and the model co-evolve during training. A pre-RL phase uses the base model's own rollouts to pre-retire low-fitness skills, seeding supervised fine-tuning; RL then continues with selective retirement, stabilization and LLM-guided mutation each iteration. The authors report the highest aggregate success rate across several interactive agent benchmarks, up to 7.8% relative improvement over the strongest baseline while keeping the library compact, and release SkillFurnace, a 5k+ record annotated dataset of retirement-filtered SFT trajectories, evolved skill libraries and retirement events.
Hugging Face · Papers · 🔥 0
Experience and discussion from the community
Share my arXiv experienceAsk about arXiv
Nobody has shared their experience with arXiv yet.