xClean Tools

AI Papers — Daily Top 5

UPDATED 2026-09-05 13:05 PDT

2026-08-03 · MONDAY · 13:05 PDT

  1. 1 Mental World Modeling 52 UPVOTES · HAO FEI ET AL. · MENTAL WORLD MODEL · ARXIV 2607.27201 The paper proposes Mental World Modeling, a framework that makes hidden mental states such as beliefs, desires, intentions, and feelings core variables of a world model. The authors argue that a model tracking only the physical scene predicts the wrong action for the right-looking scene, since human behavior is driven by what agents know and believe rather than by the scene alone.
  2. 2 N_0-TWAM: Scaling Tactile-Native World-Action Model for Contact-Rich Manipulation 26 UPVOTES · NEOTEAI TEAM ET AL. · NEOTEAI · ARXIV 2607.23783 NeoteAI presents N_0-TWAM, a tactile-native world-action model for contact-rich manipulation that predicts both future vision and future contact. It is pre-trained with visuo-tactile joint training on demonstrations spanning six embodiments and 450 tasks, using the force-based NeoForce representation to condition action generation. The authors call it the first such model trained at large scale.
  3. 3 SAF-OPD: Stable Advantage Fusion for On-Policy Distillation 23 UPVOTES · YIFAN DING ET AL. · ARXIV 2607.29209 The paper combines RL with verifiable rewards, which spreads one sparse reward across all tokens, and on-policy distillation, which scores each token against a stronger teacher but caps performance at teacher quality. The authors show that fusing the two advantages with a fixed coefficient triggers entropy collapse, and propose Stable Advantage Fusion to fix the underlying miscalibrations.
  4. 4 Scaling Properties of Text Conditioning in Visual Generation 23 UPVOTES · ZILONG CHEN ET AL. · BYTEDANCE SEED · ARXIV 2607.29679 ByteDance Seed researchers measure how text conditioning scales in visual generation, a question rarely studied because diffusion loss does not scale with prompt token count. They find that converged diffusion loss scales with the amount of structured language in the prompt, decreasing approximately linearly with a likelihood-based measure of prompt structure across controlled training runs.
  5. 5 Fewer Clarifications, Better Code: Benchmarking Cross-Session Personalized Ambiguity Adaptation in Coding Assistants 19 UPVOTES · ZIJIAN XU ET AL. · HKUST · ARXIV 2607.26611 HKUST researchers formulate personalized ambiguity adaptation as a new task for coding assistants: using a user's resolved session history as memory to disambiguate recurring, user-specific ambiguities in new sessions. Existing methods handle each ambiguous request in isolation, typically by asking clarifying questions, and the benchmark tests whether assistants can instead learn a user's patterns across sessions.

2026-08-01 · SATURDAY · 13:03 PDT

  1. 1 MPIE-Bench: Benchmarking Anatomically Plausible Multi-Person Interaction Editing 37 UPVOTES · JIAJIA LIN ET AL. · MUSET.AI · ARXIV 2607.27616 Researchers introduce MPIE-Bench, a 2,500-sample benchmark for editing multiple named people into shared contact actions such as embracing or carrying. Built from video-mined editing triplets across 405 scenes and 14 interaction categories, it targets failures like fused limbs and interpenetrating bodies that VLM-judge checklists miss but humans spot easily.
  2. 2 ACE-Data-0: Human-Centric Ambient Capture as Embodied Data Engine 34 UPVOTES · YUKANG CAO ET AL. · ARXIV 2607.28625 The Ambient Capture Engine (ACE) turns real home environments into spatially calibrated, temporally synchronized recording studios for embodied AI data. It addresses the field's data bottleneck by capturing first-person perception, whole-body motion, dexterous manipulation, object state, sound, and touch together as humans pursue goals over time.
  3. 3 Beyond Borrowed Histories: Person-Aligned User Simulation for Interactive Role-Playing Evaluation 30 UPVOTES · YUHANG ZHU ET AL. · MUSET.AI · ARXIV 2607.27816 This paper targets evaluation of role-playing agents, one of the biggest consumer uses of LLMs. Arguing that continuing a fixed dialogue history under a fixed rubric misjudges systems, the authors propose person-aligned user simulation so agents are scored in interactive conversations shaped by the simulated user rather than borrowed transcripts.
  4. 4 RefCaptioner: Multi-Reference Image-Grounded Video Captioning 25 UPVOTES · TENGFEI LIU ET AL. · KLING TEAM · ARXIV 2607.28509 The Kling team introduces multi-reference image-grounded video captioning, a task requiring factual video descriptions with phrase-level grounding to reference images. RefCaptioner, a two-stage post-training framework combining mixed-data SFT with Hierarchical Coverage-Discounted GRPO, improves reference selection, phrase binding, and distractor rejection.
  5. 5 See2Think: Do Multimodal Models Really Use Intermediate Visual States? 22 UPVOTES · SIYU YAN ET AL. · ARXIV 2607.26769 See2Think asks whether multimodal models that sketch, annotate, and generate intermediate images during reasoning actually rely on those visual states. The framework pairs See2ThinkBench, 1,200 open-ended visual tasks, with a Visual Action-of-Thought protocol that diagnoses how intermediate visual states are generated, rendered, and used rather than only scoring final answers.

2026-07-31 · FRIDAY · 13:05 PDT

  1. 1 Qwen-UI-Agent Technical Report: Toward Next-Generation Real-World Centric Foundation GUI Agents 264 UPVOTES · HANZHANG ZHOU ET AL. · TONGYILAB · ARXIV 2607.28227 TongyiLab presents Qwen-UI-Agent, a real-world centric foundation GUI agent spanning mobile, computer-use, web, and DeepSearch environments. The report targets agents that operate reliably on real devices, combine GUI interaction with CLI execution, and complete long-horizon tasks while improving with minimal human effort. It leads the day's Hugging Face papers with 264 upvotes.
  2. 2 Metis: Memory Foundation Model 213 UPVOTES · ZEYU ZHANG ET AL. · MEMTENSOR · ARXIV 2607.26760 MemTensor introduces memory foundation models, which build native memory capability directly into foundation models instead of relying on external memory modules. The work formalizes native memory in part as a persistent, dynamically evolving memory state, taking a first step toward agents that remember by design rather than by bolted-on retrieval.
  3. 3 AskChem: Claim-Centered Infrastructure for Chemistry Literature Synthesis 199 UPVOTES · BING YAN ET AL. · NEW YORK UNIVERSITY · ARXIV 2607.28618 NYU researchers present AskChem, a claim-centered infrastructure for cross-paper chemistry search. Instead of returning ranked document lists, it converts each paper into atomic, typed claims grounded by a source DOI, so scientists and AI agents can assemble cross-paper answers with verifiable provenance.
  4. 4 PhiZero: A World Model Built Around Physical Language 148 UPVOTES · SHUYAO SHANG ET AL. · ARXIV 2607.28624 PhiZero is a physical world model built around physical language, a compact discrete representation of world-state transitions learned from in-the-wild videos via self-supervision. Rather than predicting future video directly in pixel space, it uses this representation to reason explicitly about how the physical world evolves.
  5. 5 Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering 121 UPVOTES · JUNLIN YANG ET AL. · FRONTIS AI · ARXIV 2607.28568 Frontis AI post-trains Frontis-MA1, a 35B meta-evolution agent for machine learning engineering, on OpenMLE, an open full-stack system for recursive self-improvement research. The stack spans verifiable task environments with execution feedback, operator learning, and long-horizon search, treating MLE as a concrete testbed for AI that improves the process of building AI.

2026-07-30 · THURSDAY · 12:36 PDT

  1. 1 TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAM 118 UPVOTES · HENGYI XIE ET AL. · H-EMBODVIS · ARXIV 2607.27205 Vision-language-action models usually route what the robot sees through a language model before decoding an action, which costs compute and memory on every policy call. TurboVLA reworks that pathway, and the authors report control running at 32 Hz on a single RTX 4090 inside 1 GB of VRAM.
  2. 2 CoRT: Counterfactual Replay for Token-Level Rubric-Guided Policy Optimization 73 UPVOTES · BO-WEN ZHANG ET AL. · BYTEDANCE · ARXIV 2607.25659 Rubric-based reinforcement learning grades model outputs against explicit written criteria, but GRPO-style pipelines flatten those judgments into a single response-level number. CoRT uses counterfactual replay to push the rubric signal down to individual tokens.
  3. 3 HumanCLAW: Can Vision-Language Models Act Through a Body? 65 UPVOTES · SIYAO LI ET AL. · META RESEARCH · ARXIV 2607.27180 Judging whether a vision-language model can act through a physical body is hard because the outcome mixes the model's decision with motor control, so a failed task does not say which part went wrong. HumanCLAW is built to separate the two.
  4. 4 DecoEvo: Score-Decoupled Co-Evolution of Solver and Rubric-Generator 54 UPVOTES · JIANGWANG CHEN ET AL. · QWEN BUSINESS UNIT · ARXIV 2607.25675 Text-space optimization adapts a language model by editing external natural-language artifacts rather than weights, which keeps what changed readable and lets the model stay a black box. DecoEvo evolves the solver and the rubric generator together, with their scores decoupled.
  5. 5 CLBench-V: Evaluating Multimodal Context Learning from Grounding to Knowledge 40 UPVOTES · LAI WEI ET AL. · ARXIV 2607.25294 Real tasks often require a model to learn from the context it is given rather than lean only on what it absorbed in pre-training. CLBench-V evaluates that ability across multimodal settings, spanning visual grounding through to knowledge-level questions.

2026-07-29 · WEDNESDAY · 12:40 PDT

  1. 1 HiFi-UMI: Learning Deployable Manipulation Policies from High-Fidelity UMI Data Alone 126 UPVOTES · SIMPLE AI ET AL. · SIMPLE WORLD LAB · ARXIV 2607.25895 Deployable manipulation policies are held back by data that is either accurate or scalable, rarely both: teleoperation is precise but costly, while robot-free UMI capture scales and is usually relegated to pre-training with a small real-robot anchor added later. The authors ask whether raising UMI fidelity can remove that anchor entirely.
  2. 2 A New Role for Relevance: Guiding Corpus Interaction in Agentic Search 79 UPVOTES · JIANGNAN LI ET AL. · TENCENT · ARXIV 2607.24223 Retrieval agents use relevance to pick top-k content, but the authors argue relevance alone cannot localize, compose, or verify the evidence a complex question needs. Direct Corpus Interaction supports those finer operations through grep-style exploration, yet its relevance-agnostic search surfaces useful clues late and delays convergence.
  3. 3 StateAct: Program State, before Pixels, for Long-Horizon Computer-Use Agents 55 UPVOTES · YAN YANG ET AL. · SALESFORCE AI RESEARCH · ARXIV 2607.22798 Computer-use agents are usually improved by sharpening perception of screenshots. The authors point out that a screenshot is a lossy rendering of the underlying program state - files, application backends, the DOM - and that different states can render to identical pixels, so they condition the agent on that state instead.
  4. 4 ReDesign: Recovering Editable Design Structures from Images via Agentic Decomposition 54 UPVOTES · JOOYEOL YUN ET AL. · KAIST AI · ARXIV 2607.25565 Turning a raster image back into an editable design file is a costly bottleneck, because editability depends on recovering typography, vector geometry, colors, grouping, and layer ordering together. ReDesign is an agentic framework that grows an editable layer hierarchy by selecting and composing specialized tools across those modalities.
  5. 5 CodeNib: A Multi-View Data System for Serving Repository Context to Coding Agents 34 UPVOTES · ZHONGMING YU ET AL. · SYSEVOL AI RESEARCH · ARXIV 2607.25431 Coding agents keep rediscovering the same repository context because indexes, language servers, and task-local histories are disconnected. CodeNib builds lexical, dense, and structural views per commit, maps results to repository-relative source ranges, and serves ranked search, symbol navigation, and bounded context from one runtime.

2026-07-28 · TUESDAY · 12:40 PDT

  1. 1 Kimi K3: Open Frontier Intelligence 250 UPVOTES · KIMI TEAM ET AL. · MOONSHOT AI · ARXIV 2607.24653 The report introduces Kimi K3, a 2.8T parameter mixture-of-experts model with 104 billion activated parameters, native vision, and a 1-million-token context window. Alongside Kimi Delta Attention and Attention Residuals, a Stable LatentMoE design activates 16 of 896 routed experts per token, which the team credits for roughly 2.5x better scaling efficiency.
  2. 2 JarvisHub: An Open Harness for Canvas-Native Multimodal Creative Agents 101 UPVOTES · YUNLONG LIN ET AL. · ARXIV 2607.23588 Creative AI is moving from single-step asset generation toward long-horizon production, the authors argue. Real creative work involves references, drafts, alternatives, edits, failed attempts, version relations, tool actions, and human feedback, which together form an evolving project state that isolated prompt-output exchanges cannot hold.
  3. 3 From Proprietary to Open-Source: Bridging the Distribution Gap via Multi-Agent Protocol Distillation in Agentic Search 64 UPVOTES · JUNLIN LIU ET AL. · ARXIV 2607.24280 Agentic search interleaves multi-step reasoning with retrieval, but outcome-based reinforcement learning gives it only sparse supervision. Proprietary models make strong teachers for denser guidance, yet their hidden logits rule out conventional logit-matching, so the authors distill at the level of a multi-agent protocol instead.
  4. 4 Rethinking Classifier-Free Guidance in On-Policy Diffusion Distillation 61 UPVOTES · BINGNAN LI ET AL. · ARXIV 2607.24731 On-policy distillation adapts a diffusion model by querying a teacher along the student's own trajectories, but its behaviour under classifier-free guidance is poorly understood. The authors show that matching guided velocities directly is under-identified at the branch level, since positive- and negative-branch errors can offset each other.
  5. 5 Progress Reward Modeling for Robotic Learning: A Comprehensive Survey 55 UPVOTES · JIANSHU ZHANG ET AL. · NORTHWESTERN UNIVERSITY · ARXIV 2607.21655 A survey of reward models that score progress rather than completion. A terminal success signal only says whether a task finished, not whether behaviour is advancing, stalling, or undoing earlier progress. The authors note the literature still lacks a shared framework, with methods differing in observations and goal specification.

2026-07-27 · MONDAY · 00:40 PDT

  1. 1 Skill Self-Play: Pushing the Frontier of LLM Capability with Co-Evolving Skills 16 UPVOTES · SIYUAN HUANG ET AL. · QWENBUSINESSUNIT-EDU · ARXIV 2607.22529 The paper targets the dilemma in self-evolving LLM training: environment-bound methods get precise feedback but stay narrow, while open-ended self-generation broadens tasks and loses reliable verification. It proposes agent skills as the middle ground, since each skill gives deep and verifiable execution in its own domain.
  2. 2 DataPrep-Bench: Benchmarking LLMs as Training Data Preparators 14 UPVOTES · HAO LIANG ET AL. · ARXIV 2607.20465 A benchmark for how well LLMs, agents, and data-centric workflows prepare training data end to end. It splits the job into data construction, which turns raw sources into supervised data, and data quality evaluation, which predicts a candidate dataset's downstream training value before training runs.
  3. 3 Molt: A Scalable PyTorch-Native Training Framework for Agentic Reinforcement Learning 10 UPVOTES · JIAN HU ET AL. · NVIDIA · ARXIV 2607.21653 NVIDIA researchers argue that agentic RL work means constant algorithm changes, and that mainstream frameworks make each change thread through trainer, distributed backend, and rollout glue. Molt keeps the codebase small enough for a researcher, or an AI coding assistant, to hold in full and trace end to end.
  4. 4 Scaling Native Multimodal Pre-Training From Scratch 7 UPVOTES · HAOYUAN WU ET AL. · TENCENT HUNYUAN · ARXIV 2607.22043 Native multimodal pre-training trains from scratch on multimodal inputs, avoiding the text-only limits and late-fusion asymmetries of standard LLMs. The scaling properties of that paradigm are still uncharacterized, so this work investigates the optimal model size and token allocation for it.
  5. 5 LAMAR: An Open Language-Aware Multilingual Alignment Reranker 6 UPVOTES · SEONGTAE HONG ET AL. · NLP & AI - KOREA UNIVERSITY · ARXIV 2607.22042 In multilingual RAG a retriever returns documents in several languages, which are then reranked. The authors show existing rerankers do not consistently prefer documents in the query's own language when semantically equivalent ones exist, even though document language affects the generated answer, and they release LAMAR in response.