xClean Tools

AI Papers — Daily Top 5

UPDATED 2026-09-05 13:05 PDT

2026-08-11 · TUESDAY · 01:04 PDT

  1. 1 SWE-Bench ProMax: Benchmarking Agents on Large-Scale Multilingual Code Refactoring 70 UPVOTES · YULING SHI ET AL. · BYTEDANCE · ARXIV 2608.09802 A ByteDance team proposes SWE-Bench ProMax, a coding-agent benchmark built on large-scale multilingual refactoring instead of bug fixes. The authors cite an audit finding that nearly 60 percent of unsolved SWE-bench Verified instances have flawed tests, and that frontier models can reproduce gold patches verbatim. Refactoring demands behaviour-preserving edits coordinated across many files.
  2. 2 SFT Conflicts, RL Coexists: A Theoretical and Empirical Analysis of Multi-Task Learning for LLMs 40 UPVOTES · KEJIAN ZHU ET AL. · CHINESE ACADEMIC OF SCIENCE INSTITUTE OF AUTOMATION · ARXIV 2608.03573 Researchers at the Chinese Academy of Sciences Institute of Automation report that supervised fine-tuning suffers severe task conflicts under multi-stage training while reinforcement learning lets diverse tasks coexist. Traced to the parameter level, RL induces sparse and approximately orthogonal updates across tasks. The paper offers a theoretical account based on multi-task gradient interference.
  3. 3 Beyond Simply Environment Scaling: Designing Effective Environment Distributions for Multimodal Agent Learning 39 UPVOTES · KEJIAN ZHU ET AL. · CHINESE ACADEMIC OF SCIENCE INSTITUTE OF AUTOMATION · ARXIV 2608.03571 Simply enlarging the pool of multimodal training environments does not always help agents, this group finds. They instead shape the distribution along two axes: Ability-aware Environment Selection picks diverse environment sets, while a difficulty structure controls how hard those environments are. The work argues distribution design matters more than raw environment count.
  4. 4 Macaron-V1: Towards Open Continual Learning with Self-Improvement and Mixture-of-LoRA 38 UPVOTES · MIND LAB ET AL. · MIND LAB · ARXIV 2608.09819 Mind Lab releases Macaron-V1, an open agent-model family designed to keep learning from real environments after deployment. Adaptation works by recursively improving versioned model-harness pairs, with each configuration evaluated under an external contract to construct its successor. A Mixture-of-LoRA architecture freezes the base model and selects one specialist adapter per user turn.
  5. 5 Motif 3: Technical Report 23 UPVOTES · JUNGHWAN LIM ET AL. · MOTIF TECHNOLOGIES · ARXIV 2608.09119 Motif Technologies details Motif 3, a decoder-only Mixture-of-Experts model with 314 billion total parameters and 13.2 billion active per token. Each sparse layer holds 384 routed experts with eight selected per token, giving large expert capacity at limited compute. The architecture centres on Grouped Differential Latent Attention, which pairs grouped differential attention with compressed key-value representations.

2026-08-10 · MONDAY · 13:03 PDT

  1. 1 StreamArena: Toward Continuous, Interactive, and Long-Horizon Agentic Streaming Video Understanding 14 UPVOTES · XICHEN ZHANG ET AL. · XIAOHONGSHU · ARXIV 2608.05703 StreamArena is a benchmark for hour-scale interactive video understanding, built from 243 full-length videos averaging 88.8 minutes. The authors argue that current evaluations rely on short clips and multiple-choice formats, which let a baseline reading only the last four frames match far more complex streaming models.
  2. 2 DCAS: Decoupling CLI Agent Scaffolding to Internalize Planning across Scaffolds 12 UPVOTES · KISHANTHAN THANGARAJAH ET AL. · CENTRE FOR SOFTWARE EXCELLENCE · ARXIV 2608.06113 Open trajectory datasets for command line software engineering agents are collected almost entirely under one scaffold, OpenHands. The authors show that models fine-tuned on that data score well there but degrade under any other scaffold, while untrained base models do not, and trace the gap to learned planning structure.
  3. 3 When Activation Oracles Learn Not to Read: Concept-Specific Blind Spots in Fine-Tuned Oracles 10 UPVOTES · TOBIAS BERSIA ET AL. · QUEEN MARY UNIVERSITY OF LONDON · ARXIV 2607.23379 Activation oracles are language models trained to answer questions about another model's internal activations. This study argues they are learned systems rather than neutral readouts, and uses a controlled taboo word guessing setup to show concept-specific blind spots shaped by training data and reporting behaviour.
  4. 4 Efficient Knowledge Distillation for LLMs: Offline Top-K Logits and a Fused Chunked KL Loss 9 UPVOTES · BAKBERGEN RYSKULOV ET AL. · MULTIVERSE COMPUTING · ARXIV 2608.03796 Compressed small models are usually recovered through knowledge distillation, an expensive step that largely decides final quality. This practitioner study reports that offline distillation, caching a teacher's top-K logits once and training the student against that cache, matches online distillation at near-identical quality.
  5. 5 Douyin Multimodal Embedding Model Technical Report 7 UPVOTES · HAONAN CHEN ET AL. · BYTEDANCE · ARXIV 2608.02148 ByteDance describes the multimodal embedding model behind search and recommendation on Douyin. The report frames the industrial constraint as needing both efficiency at billion-scale indexing and fine-grained discrimination, which it says contrastive models and chain-of-thought based models each satisfy only in part.

2026-08-08 · SATURDAY · 13:04 PDT

  1. 1 AgentOPSD: Recursive Self-Distillation for Agentic Reinforcement Learning 75 UPVOTES · ZI-HAN WANG ET AL. · TSINGHUA UNIVERSITY · ARXIV 2608.05987 A Tsinghua-led team proposes AgentOPSD, a critic-free method for turn-level credit assignment in agentic reinforcement learning. It aggregates token-level log-probability gaps between a privileged teacher and the student policy, then applies that signal recursively so the few pivotal decisions in long-horizon, multi-turn tasks are credited. It is the day's most upvoted paper with 75 votes.
  2. 2 WorldClaw: Agentic 3D Open-World Generation at Scale 50 UPVOTES · CHUNCHAO GUO ET AL. · TENCENT HUNYUAN · ARXIV 2608.05248 Tencent Hunyuan presents WorldClaw, a fully agentic coarse-to-fine framework that builds explorable 3D worlds from open-ended text. Planning agents turn a prompt into a structured specification of regions, terrain, assets, materials and spatial relations, which the system then realizes as a coherent scene with explicit assets suitable for editing and reuse.
  3. 3 ChronoVision: Temporal Reasoning via Latent State Reconstruction 33 UPVOTES · YIFAN SHEN ET AL. · PEDIAMED AI · ARXIV 2608.05631 ChronoVision targets a known weakness of multimodal LLMs: strong passive perception but poor multi-step temporal reasoning, which the authors trace to the ambiguity of describing continuous visual change in words. During supervised fine-tuning, a reconstructive visual head predicts the latent representation of the final transformed state while an ROI attention module locates the regions that change.
  4. 4 Learning from Failures: Retrieval-Centric CoT via Hard Negatives for Unified Multimodal Retrieval 31 UPVOTES · ZELONG SUN ET AL. · DEEPGLINT · ARXIV 2608.06060 DeepGlint researchers build chain-of-thought rationales for multimodal retrieval out of hard negatives rather than the query alone. Existing CoT retrievers explain what the query describes, which leaves vision-language models confusing semantically similar candidates; grounding the reasoning in near-misses surfaces the fine-grained cues that separate a true target from them.
  5. 5 From Economic Agents to Agentic Economies: A Systems Blueprint for Economic World Models 28 UPVOTES · JIALE HAN ET AL. · FREEDOMAI · ARXIV 2608.06020 A team from FreedomAI sets out an implementation roadmap for economic world models, generative systems that simulate an economy from the inside by modeling heterogeneous agents, their beliefs and actions, and the markets and institutions they interact through. The paper organizes such systems into a six-level capability ladder starting from fixed rule-based agents.

2026-08-07 · FRIDAY · 13:06 PDT

  1. 1 OSReward: Instituting Standardized Evaluation for Cross-Platform Computer-Use Reward Models 49 UPVOTES · QIUSHI SUN ET AL. · NLP GROUP OF THE UNIVERSITY OF HONG KONG · ARXIV 2607.28609 A 23-author team led from the University of Hong Kong introduces OSReward, a standardized evaluation for the vision-language models used to judge computer-use agent trajectories. The field increasingly relies on VLM judges because neither hand-written verifiers nor human annotators scale, yet whether those judges are reliable has gone largely unexamined.
  2. 2 Interpretable MEG Decoding of Perceived Speech: Cortical Sources and the Stimulus Features That Drive Retrieval 36 UPVOTES · ILIA SEMENKOV ET AL. · ARXIV 2608.01481 Deep networks can retrieve short segments of heard speech from non-invasive MEG recordings, but their weights do not correspond to any electrophysiological quantity. This paper redesigns both ends of a high-performing retrieval architecture, replacing spatial attention over a flattened sensor layout with spherical harmonics defined on the three-dimensional helmet geometry.
  3. 3 GST-Bench: Can VLMs Develop Global Spatial Awareness from Video? 30 UPVOTES · QIFENG ZHANG ET AL. · BYTEDANCE SEED · ARXIV 2608.05747 ByteDance Seed releases GST-Bench, a video question-answering benchmark for global spatial awareness rather than the local, few-viewpoint perception that existing tests measure. Its human-verified questions are drawn from 6,790 minutes of synthetically generated video and require models to reason about viewpoints never shown in the input.
  4. 4 OneDayAgent: Towards a Long-Horizon Harness for Autonomous Agents 29 UPVOTES · JINGSHENG ZHENG ET AL. · ZJUNLP · ARXIV 2608.05013 OneDayAgent is a harness for open-ended everyday requests that run long, cross several environments, and mix modalities. Prior work has attacked goal drift, state loss, and context overflow one at a time; the Zhejiang University group asks whether a single harness can handle all three at once and stay effective across different model backends.
  5. 5 EnvACE: Internalizing Environment Dynamics via World Rehearsal for Agentic Reinforcement Learning 28 UPVOTES · ZISHAN XU ET AL. · TENCENT · ARXIV 2608.06197 Tencent's EnvACE trains tool-using agents without touching a real or simulated environment. The policy alternates between issuing a tool call and playing the environment that answers it, a loop the authors call world rehearsal, then conditions later decisions on its own rehearsed responses. The target is the cost of building and verifying executable training environments.

2026-08-06 · THURSDAY · 13:07 PDT

  1. 1 ABSeeker: Training Long-Horizon Search Agents via Answer-Backtracked Credit Assignment 52 UPVOTES · YIJUN LU ET AL. · SHANGHAI JIAO TONG UNIVERSITY · ARXIV 2608.05102 Researchers at Shanghai Jiao Tong University propose Answer-Backtracked Credit Assignment (ABC), a fine-grained framework for training long-horizon search agents. Rather than treating all steps in a trajectory uniformly during SFT and RL, ABC converts sparse trajectory-level outcomes into per-step credit, separating useful actions from erroneous or redundant ones. It drew 52 upvotes, the day's highest.
  2. 2 Towards Physics of Multimodal Pretraining: Knowledge Flow, Modality Synergy, Early Unification, and Recipes 39 UPVOTES · JUNLIN HAN ET AL. · AI AT META · ARXIV 2608.05000 Meta researchers run a systematic empirical exploration of natively unified multimodal pretraining. Controlled experiments on synthetic and large-scale real-world datasets yield four insights spanning knowledge flow, modality synergy, early unification, and training recipes, aiming to clarify a design space that unified training has so far left underexplored.
  3. 3 Quo Vadis, World Modeling? 31 UPVOTES · YU YANG ET AL. · SHANGHAI AI LABORATORY · ARXIV 2608.02713 A 20-author team centered at Shanghai AI Laboratory takes stock of world modeling for continually improving agents. Arguing that classical future-state prediction is useful but narrow, the paper conceptualizes agent-centric interactive world models that give agents cheaper, more controllable feedback than direct real-environment interaction.
  4. 4 PAST-Bench: Benchmarking the Foundations of Recursive Self-Improvement in Personal Agents 28 UPVOTES · SHUHAN XUE ET AL. · PRINCETON UNIVERSITY · ARXIV 2608.04003 Princeton researchers introduce PAST-Bench, a benchmark for whether personal AI agents actually improve from retained experience. Agents run ordered sequences of fresh-session tasks under matched conditions that switch retained experience on and off, spanning 26 scenarios, isolating the foundations of recursive self-improvement.
  5. 5 LLaDA MoE v2: Scaling Mixture-of-Experts Diffusion Language Models 27 UPVOTES · FENGQI ZHU ET AL. · GSAI-ML · ARXIV 2608.03457 LLaDA MoE v2 characterizes how optimization hyperparameters, compute allocation, and architecture scale for mixture-of-experts diffusion language models. The study identifies quantitative differences from autoregressive scaling trends, including an optimal batch size that grows faster and an optimal learning rate that decays more rapidly with compute.

2026-08-05 · WEDNESDAY · 13:07 PDT

  1. 1 MerchantBench: Benchmarking LLM Agents for Long-Term Coherence in E-Commerce Operations 83 UPVOTES · QIMING SHI ET AL. · ALIBABA · ARXIV 2607.28956 Alibaba researchers present MerchantBench, a benchmark testing whether LLM agents stay coherent over long horizons in seller-side e-commerce operations. Agents run a persistent store where actions constrain future choices and feedback arrives at varying delays, so incoherent decisions accumulate into measurable costs. It targets a gap left by benchmarks built on bounded tasks with immediate success criteria.
  2. 2 AURORA-LM: Autoencoding Unified Representation for Continuous-Latent Diffusion Language Modeling 71 UPVOTES · JIAJUN LIANG ET AL. · NANJING UNIVERSITY · ARXIV 2608.02602 Nanjing University researchers propose AURORA-LM, which generates text with diffusion in a continuous latent space rather than discrete tokens. Instead of compressing latents to ease diffusion, it keeps a high-capacity, decodable text latent and trains the diffusion model to learn its distribution directly, preserving token-level fidelity. The approach brings text closer to how images, video, and audio are modeled.
  3. 3 Hunyuan3D-Buffalo 1.0: A Unified Multimodal Model for Scalable 3D Generation, Understanding, and Editing 70 UPVOTES · JUNLIANG YE ET AL. · TENCENT HUNYUAN · ARXIV 2608.02711 Tencent Hunyuan presents Hunyuan3D-Buffalo 1.0, a single architecture unifying 3D understanding, text-to-3D generation, instruction-guided 3D editing, and text-grounded part generation. To overcome the shortage of geometrically consistent editing data, the team built an 87-million-scale 3D multimodal dataset for training. It extends the unified-model trend from images into 3D content creation.
  4. 4 VAD: Attributing Visual Evidence for Target Reconstruction in Multimodal On-Policy Distillation 43 UPVOTES · KANGNING ZHANG ET AL. · ARXIV 2607.28590 This paper introduces Visual Attribution Distillation for multimodal on-policy distillation, where a privileged teacher corrects student-generated trajectories. Because the teacher's next-token corrections mix visual signals with linguistic priors and teacher-specific effects, VAD uses counterfactual target reconstruction to estimate which part of each correction is actually supported by visual evidence.
  5. 5 Video-DeepResearch: Towards the Next-Generation Multimodal Deepresearch Agent 43 UPVOTES · ZHEN FANG ET AL. · ARXIV 2608.03979 Video-DeepResearch extends multimodal research agents from static images to continuous video streams, pairing dense spatiotemporal grounding with open-web exploration. The authors identify two bottlenecks in current models, a bias toward textual search over visual tools and reliance on parametric memory instead of genuine tool use, and address them with a decoupled perception-exploration pipeline.

2026-08-04 · TUESDAY · 13:04 PDT

  1. 1 DAPD: Dual-Anchored Policy Distillation 54 UPVOTES · JIANYU WU ET AL. · SHANGHAI AI LABORATORY · ARXIV 2608.01735 Researchers at Shanghai AI Laboratory identify a privilege illusion in on-policy distillation: students trained against privileged teachers learn behavior they cannot reproduce from their inference-time context. The paper traces the failure to information asymmetry between teacher and student and proposes Dual-Anchored Policy Distillation to resolve it.
  2. 2 Progressive Agent Skill Generation via Reinforcement Learning 49 UPVOTES · JUNHAO SHEN ET AL. · THE CHINESE UNIVERSITY OF HONG KONG - DATABASE GROUP · ARXIV 2608.01678 Skill-alpha frames agent skill generation as a reinforcement learning problem instead of relying on hand-designed heuristics or pipeline-style consolidation. Because skills lack a natural supervision signal for relevance or correctness, the method scores them by whether they improve the agent's behavior on downstream tasks.
  3. 3 Meshy T2: Fast Native Mesh Generation with Flow Matching 47 UPVOTES · JIALE XU ET AL. · MESHY · ARXIV 2607.28675 Meshy T2 is a native mesh generation framework built on flow matching, aimed at producing artist-style topology fast enough for interactive 3D asset creation. At its core, a vertex-set mesh VAE encodes a mesh into one continuous latent token per vertex, sidestepping the slow, error-prone autoregressive token decoding of mainstream approaches.
  4. 4 UEmbed: Unified Sparse and Dense Multimodal Embeddings 41 UPVOTES · TINGYU SONG ET AL. · ALIBABA-NLP · ARXIV 2608.02583 UEmbed is a decoder-only multimodal embedding model from Alibaba-NLP that produces both sparse lexical and dense representations in a single causal forward pass. It extends learned sparse retrieval beyond encoder-style bidirectional architectures and drops the auxiliary cross-modal modules earlier multimodal systems leaned on.
  5. 5 AISPA: User-Centric System Prompt Auditing for Large Language Model Applications 34 UPVOTES · XIANGNING LIN ET AL. · STANFORD UNIVERSITY · ARXIV 2607.28617 Stanford researchers introduce AISPA, a user-centric framework for systematically auditing the system prompts that govern commercial AI applications. It evaluates parts of a prompt along eight dimensions that matter to users, targeting the trust and accountability gap created by prompts that are rarely disclosed to the public or regulators.