xClean Tools

AI Papers — Daily Top 5

UPDATED 2026-11-02 13:14 PST

2026-11-02 · MONDAY · 13:14 PST

  1. 1 REA: A Routed Embodied Agent Runtime for Asynchronous Brain-Cerebellum Control 43 UPVOTES · PENGFEI ZHOU ET AL. · ARXIV 2610.42053 REA is a runtime that lets a slow reasoning model, a router and fast reactive policies share one embodied control loop under the rule that no running policy waits on the model. It records each policy's measured success in a route memory and consults the model only at task boundaries. On 60 LIBERO-Pro tasks it reports 95.9 percent success against 94.7 for the best single agent, with 25 percent fewer control steps.
  2. 2 Apriel Voice 27B Preview: An Open Model for Speech, Reasoning, and Tool Use 32 UPVOTES · AMAN TIWARI ET AL. · SERVICENOW-AI · ARXIV 2610.40738 ServiceNow describes Apriel Voice 27B Preview, an open speech-language model that takes speech in and emits both replies and tool calls from one backbone, with no separate dialogue model. It is built on a post-trained Gemma 4 26B-A4B checkpoint and adapted to speech in two stages with online distillation. In thinking mode it reports 98.90 on Big Bench Audio and 1.91 seconds mean latency.
  3. 3 WeFair: Fidelity-Aware Alignment from Human Feedback for Personalized Query Rewriting in Short-Video Search 26 UPVOTES · YINGTAO REN ET AL. · TENCENT · ARXIV 2610.16927 Tencent trains a compact query rewriter for WeChat short-video search by grading human feedback on how directly it reveals personalized intent, across demonstration, preference and verification stages. The paper argues that optimizing for what users watch next drifts from what they asked for. An online A/B test reports a 2.593 percent relative lift in effective video views and 2.350 percent fewer reformulations.
  4. 4 Latent Reasoning in Continuous Space for Unified Multimodal Models 24 UPVOTES · BYUNGWOO JEON ET AL. · KAIST AI · ARXIV 2610.42149 KAIST AI proposes LARC, which moves chain-of-thought reasoning for image generation out of discrete tokens and into continuous hidden states. The authors blame that modality mismatch for unified models failing to inherit test-time scaling, and train in two stages: curriculum fine-tuning, then reinforcement learning on information gain without ground truth. LARC reports a 7 point GenEval gain over BAGEL.
  5. 5 OSWAM: An Effective World Action Model with One-Step Action Generation 22 UPVOTES · GUOWEN ZHANG ET AL. · THE HONG KONG POLYTECHNIC UNIVERSITY · ARXIV 2610.35850 OSWAM cuts world action models for driving and embodied control down to a single diffusion step. Instead of random noise, it seeds the process with imagination tokens that carry physical and action-conditioned priors, and adds privileged information distillation with future conditioning. The authors report quality on par with models needing tens of steps, about a 3x speedup and 40 ms latency on an RTX 5090.

2026-10-31 · SATURDAY · 13:10 PDT

  1. 1 IntReal: A Real-World Dataset and Benchmark for Multimodal Interleaved Generation 15 UPVOTES · SHAOHAN YU ET AL. · LONGCAT · ARXIV 2610.38425 IntReal is a paired dataset and benchmark for multimodal interleaved generation, where text and images alternate within a single output. The authors argue the area has stalled for two reasons: no well-defined interleaved tasks with matching training data, and no fine-grained protocol able to diagnose which capability a model lacks.
  2. 2 CoSTA: Compressed Speech Token Adaptation for Efficient Spoken Dialogue Models 12 UPVOTES · YINGZHUO DENG ET AL. · CHINESE ACADEMIC OF SCIENCE INSTITUTE OF AUTOMATION · ARXIV 2610.21143 CoSTA is a post-hoc framework that separates the generation and synthesis token spaces of spoken dialogue models while keeping the original speech tokenizer and decoder in place. Such models usually share one token space, which makes autoregressive sequences long and inference expensive; CoSTA instead extends the generation space with compressed representations.
  3. 3 WebAnything: From the Wild Web to Scalable Agent Environments 12 UPVOTES · TIANCI XIE ET AL. · ARXIV 2610.37786 WebAnything automatically rebuilds executable, resettable web environments from real URLs so that web agents can be trained and evaluated reproducibly without per-site engineering. An Explorer-Builder loop reconstructs source-grounded sites, avoiding both the manual cost of human-assisted mirrors and the drift of purely generative construction.
  4. 4 Multi-Token Prediction Done Right 11 UPVOTES · JAYDEN TEOH ET AL. · MASSACHUSETTS INSTITUTE OF TECHNOLOGY · ARXIV 2610.38331 This single-author paper identifies a flaw in standard multi-token prediction training: a residual component of the MTP gradient interferes with both next-token and multi-token learning, and the interference grows with scale. The gradient is decomposed exactly into a KL self-distillation term that helps speculative decoding and the residual term that does the damage.
  5. 5 Investor's Last Exam: Can AI Agents Do Equity Investment Without Seeing the Future? 11 UPVOTES · HUANGFEI JIANG ET AL. · ARXIV 2610.38342 Investor's Last Exam is an equity-research benchmark led by experienced buy-side investors and fund managers, built around a strictly frozen information set so an agent cannot see past its decision date. It pairs a versioned financial database of 3.43 million document versions across nine source categories with professional evaluation criteria.

2026-10-30 · FRIDAY · 13:15 PDT

  1. 1 Youtu-Parsing-Omni: One Encoder, One Schema, All Modalities 38 UPVOTES · YUNHANG SHEN ET AL. · TENCENT · ARXIV 2610.39301 Tencent's Youtu-Parsing-Omni turns documents, images, audio and video into structured records using one encoder for every modality and one output schema for all ten parsing tasks. A single 5B-parameter checkpoint scores 96.96 overall on OmniDocBench v1.6 and leads open-weight models on OmniParsingBench at 75.08 average, with its margin concentrated on audio and video. Code and weights are public.
  2. 2 L3G-World: Real-Time Infinite World Generation with Latent 3D Gaussian Memory 34 UPVOTES · JINHYEOK CHOI ET AL. · KAIST AI · ARXIV 2610.40614 L3G-World keeps an interactive world model spatially consistent by holding the scene as latent 3D Gaussians, so memory grows with the extent of the scene rather than the length of the generated stream. Each Gaussian stores a VAE latent, letting the memory render straight into the diffusion transformer's latent space with no re-encoding. The authors report real-time generation at 9 frames per second.
  3. 3 TAO: Tool-Aware Joint Scheduling for LLM-Based Agentic Serving 29 UPVOTES · SHUHUAI LI ET AL. · AIOR · ARXIV 2610.37230 TAO schedules agentic serving by accounting for when external tool calls will return, rather than treating GPU inference and CPU tools as separate queues. A dynamic program places waiting sessions using tool readiness, a time-decayed cache credit per node and token footprint. Across 12 model, workload and concurrency settings it cuts latency by up to 26.8 percent and raises decode throughput by up to 29.5 percent.
  4. 4 A Broader Look into MLLM's Retrieval Heads 28 UPVOTES · JIA WANG ET AL. · TENCENT · ARXIV 2610.32411 Across 18 language, vision-language and omni-modal checkpoints, the sparse attention heads that pull evidence out of long inputs are largely shared across modalities. The shared heads are under 2 percent of all heads, yet ablating them also degrades document understanding, spatial perception and visual math. Training them with a ranking objective lifts Qwen3-VL-8B's page Recall@1 on MMDocIR by 6.03 points.
  5. 5 Proving and Improving the Platonic Representation Hypothesis with JEPAs 25 UPVOTES · RANDALL BALESTRIERO ET AL. · GALILAI-GROUP · ARXIV 2610.39302 The paper turns the Platonic Representation Hypothesis into a theorem: joint-embedding predictive architectures trained separately on any number of modalities learn the same representation up to one rotation, provided their positive pairs are two observations of a single scene. It casts next-token prediction as a JEPA with bad views, needing 42 times more samples to reach the same alignment.

2026-10-29 · THURSDAY · 13:18 PDT

  1. 1 CodeCap-Bench: Can Code Serve as a Good Video Caption? 50 UPVOTES · JIAFU TANG ET AL. · ARXIV 2610.37341 Jiafu Tang and 13 co-authors propose Code as Video Caption, where an executable program describes a video through graphical primitives, layout rules and functions of time. CodeCap-Bench then tests whether those code captions preserve as much information as natural language, which the paper argues grows long and ambiguous on fine spatial structure and temporal dynamics.
  2. 2 LociArt: A Unified Paradigm for Free-Style Local Video Stylization 45 UPVOTES · YUXUAN CHEN ET AL. · XIAOMI RESEARCH · ARXIV 2610.36456 Xiaomi Research introduces LociArt, a paradigm for local video stylization that states both where to stylize and which style to apply while leaving the rest of the frame intact. Its GRNStyle core builds on Generative Refinement Networks and adds a polar style-aware mask that embeds style identities as equal-norm directions inside the selected regions.
  3. 3 OmniShowcase: Omnidirectional Subject-Consistent Video Generation from Multi-View Images 41 UPVOTES · JUNYEONG AHN ET AL. · KAIST AI · ARXIV 2610.35684 A KAIST AI team presents OmniShowcase for generating video of a specific subject from a few reference views. Earlier subject-to-video methods condition on a single image and must invent the surfaces a moving camera later reveals; multi-view successors add information but, the paper argues, never state the 3D structure that ties those observations together.
  4. 4 RSI Arena: Towards Recursive Self-Improvement with Human Feedback 31 UPVOTES · YUE HUANG ET AL. · RSI ARENA · ARXIV 2610.33624 RSI Arena asks whether AI agents can post-train useful language models on their own. Ten frontier models each act as a training agent starting from NVIDIA Nemotron 3.5 Lightning 30B-A3B Base, with 1,000 GPU-hours and $300 of API credit over a six-day stage on a shared 64-GPU cluster, choosing their own data, procedure and checkpoints. The results are scored on 20 benchmarks and by blind human evaluation.
  5. 5 Nemotron-Labs-Diffusion-Uni: Elastic Self-Refining Unified Diffusion LLM with Cross-Paradigm On-Policy Distillation 27 UPVOTES · SHUFAN LI ET AL. · NVIDIA · ARXIV 2610.36466 NVIDIA proposes NLD-Uni, a unified discrete diffusion language model covering both visual understanding and generation. It is trained by cross-paradigm on-policy distillation, which transfers understanding from autoregressive VLMs and generation from continuous flow models into one diffusion model, and it adds a progressive self-refining diffusion process that corrupts each token gradually.

2026-10-28 · WEDNESDAY · 13:18 PDT

  1. 1 Multiscale Policy Optimization for Long-Horizon LLM Agents 316 UPVOTES · TAO REN ET AL. · IQUEST · ARXIV 2610.35013 The paper argues that long-horizon LLM agents should be optimized at the turn scale: critic values vary mainly across assistant turns, not across the tokens inside them. Multiscale Policy Optimization pairs multiscale advantage estimation with turn-level importance sampling so credit assignment and policy updates share that scale. Experiments on search, tool-use and automation tasks report consistent gains.
  2. 2 VibeVideoBench: Evaluating Multimodal Agents on Code-Driven Video Generation 143 UPVOTES · MUXI DIAO ET AL. · ARXIV 2610.35864 VibeVideoBench evaluates multimodal agents that write code to render every frame and synthesize every sound, instead of sampling from a diffusion model. Its 239 tasks in English and Chinese span 12 categories, from data stories to classroom lessons and paper explainers, each run in the agent's native harness. The premise is that diffusion video cannot reliably render an exact formula or a correct chart.
  3. 3 Memory Tax: Dissecting the Net Value of Agent Memory 99 UPVOTES · XUAN HE ET AL. · UNIVERSITY OF ILLINOIS AT URBANA-CHAMPAIGN · ARXIV 2610.34073 A large-scale evaluation across multiple benchmarks and model backbones finds that many dedicated agent memory systems underperform naive retrieval-augmented generation, leaving a negative net value. Taking RAG as the reference point, the authors split memory's net value into a memory benefit and a memory tax, both arising because the memory process changes how raw history is organized and accessed.
  4. 4 Can Discrete Tokens Serve as Semantically Lossless Representations of Visual Signals? 81 UPVOTES · JIALV ZOU ET AL. · LONGCAT · ARXIV 2610.37662 The paper revisits the widely held assumption that quantization error forces discrete visual tokens to lose semantics, capping them below continuous representations. Using multimodal understanding as a proxy for semantic fidelity, it reports that quantization error is an unreliable indicator of that loss. Discrete tokens let vision share a next-token prediction objective with language.
  5. 5 Representation Redistribution in Sparse Autoencoders 56 UPVOTES · ZIYE MEI ET AL. · OPENSAE · ARXIV 2610.21240 Sparse autoencoders decompose language model activations into sparse features, but rewriting a prompt to preserve the same request can redistribute activity across them. The authors introduce a redistribution sensitivity score that keeps only large feature responses recurring in a consistent direction, then test them on held-out prompt variants. The study covers 27 pretrained SAEs across eight language models.

2026-10-27 · TUESDAY · 13:18 PDT

  1. 1 Function-Space Temporal Self-Distillation for Policy Optimization 130 UPVOTES · ARYAN DAS ET AL. · INDIAN INSTITUTE OF TECHNOLOGY, KANPUR · ARXIV 2610.19687 A team at IIT Kanpur introduces Function-Space Policy Optimization, a single-policy method for reinforcement learning with verifiable rewards. It keeps the dense token-level signal of self-distillation, which compares an unassisted policy with a feedback-informed self-teacher, but consolidates targets in function space over time so the teacher stops drifting with the student. It leads the day's pool with 130 upvotes.
  2. 2 Work Agents: Systems, Capabilities, and Optimization 97 UPVOTES · ZISHAN XU ET AL. · SHANGHAI JIAO TONG UNIVERSITY · ARXIV 2610.35845 A survey led from Shanghai Jiao Tong University treats work agents as systems rather than models: a foundation model plus a runtime harness that executes tasks in stateful workspaces and delivers verifiable outcomes. It organizes the field along two dimensions, work persistence and work openness, to compare how reliably agents finish delegated work.
  3. 3 Continual Learning Exam: Benchmarking Long-Horizon RSI and Experiential Learning in LLM Agents 43 UPVOTES · HAOYU ZHAO ET AL. · UNIVERSITY COLLEGE LONDON · ARXIV 2610.33956 University College London introduces the Continual Learning Exam, a benchmark built with more than 50 domain experts to measure whether an LLM agent improves from accumulated experience rather than on isolated tasks. It comprises 120 human-authored tasks set in realistic professional workflows, the long-horizon dimension that current evaluations largely leave untested.
  4. 4 WorldCast: World-Consistent Character Video Generation with Large-Angle Camera Control 42 UPVOTES · LEI CAO ET AL. · ARXIV 2610.33135 WorldCast takes on character video generation along large-angle camera trajectories, where combined rotation, translation and radius changes reveal regions absent from the first frame and bring geometric distortion, appearance degradation and identity drift. The authors release CLACT, a large character-centric dataset and benchmark with accurate camera geometry, to study those coupled failures.
  5. 5 Spatial Audio Generation and Editing with Executable Plans and Self-Distilling Transfusion 38 UPVOTES · KAM MAN WU ET AL. · NOIZAI · ARXIV 2610.34415 NoizAI presents AMBIT, a framework for generating and editing native first-order Ambisonics audio. A shared Transfusion backbone couples autoregressive planning with rectified-flow rendering through executable ScenePlans that set each source's content, activity and 3D trajectory; for an edit the plan describes the whole target scene, including the sources to retain.

2026-10-26 · MONDAY · 13:16 PDT

  1. 1 Reading, Writing, Sharing, and Interacting with the World through Video 211 UPVOTES · SHENGQIONG WU ET AL. · UNIVERSITY OF OXFORD · ARXIV 2610.31668 A survey from the University of Oxford adopts world modeling as a single lens on both video foundation models and video world models. Treating an observed video as a partial, lossy and temporally ordered projection of a latent world, it organizes the literature around four complementary operations: reading, writing, sharing and interacting. It drew 211 upvotes, far ahead of the rest of the day's pool.
  2. 2 Every Weight Is a Path of Shared Discrete Decisions 18 UPVOTES · AMMAR ALI ET AL. · ITMO UNIVERSITY · ARXIV 2610.31931 ITMO University proposes TWIG, or Tree-structured Weights with Inherited Gates, a representation in which every network weight is the endpoint of a path of small discrete decisions: a sign, a shift by a power of two, a zero gate and an index into a shared table. Weights that belong together share the upper part of their paths, so each decision is stored and applied once. Any trained network can be rewritten this way.
  3. 3 Dual Exam: Controlled Evaluation of Multimodal Agents and World Models 16 UPVOTES · ZHEFAN RAO ET AL. · ARXIV 2610.31066 Dual Exam pulls apart the evaluation of multimodal agents and world models, testing agents under fixed environment responses and world models under fixed action interventions across physical, digital, social and scientific domains. The agent track compares 10 configurations on 1,088 items in 15 reporting categories; the world model track holds 400 cases across eight scenarios.
  4. 4 DirCast: Live Esports Directing and Commentary with a Single Streaming Video-Language Model 15 UPVOTES · ZIXIAN GAO ET AL. · ARXIV 2610.31734 DirCast is a single streaming video-language model that both directs and commentates a live esports broadcast, taking every player's raw camera feed and working only from what has already happened. Earlier streaming models learned to commentate, but only over a broadcast a human director had already cut. Each second DirCast scans all cameras at low resolution before deciding what goes on air.
  5. 5 TenetBench: A Clause-Anchored Russian Corpus for Evaluating Language Models against Professional Codes of Ethics 12 UPVOTES · YEGOR ANICHKOV ET AL. · ARXIV 2610.29477 TenetBench tests whether an assistant breaches a specialist's professional code, or helps them breach it, without either party noticing. It pairs 3,433 requests that conflict with clauses of fifteen Russian-language professional codes against benign twins differing only in the legitimacy of what is asked, so each case points at a specific provision rather than a broad harm category.