xClean Tools

AI Papers — Daily Top 5

UPDATED 2026-09-05 13:05 PDT

2026-08-20 · THURSDAY · 13:07 PDT

  1. 1 SemComp-Bench: Benchmarking Semantic Task Completion in Video Generation 151 UPVOTES · KEYU TU ET AL. · FRAMEX-AI · ARXIV 2608.17426 SemComp-Bench recasts video generation as an outcome-oriented task: a clip counts as a success only if it reaches the intended result and stays semantically grounded in the reference image. Evaluation judges the generated outcome rather than demanding a complete sequence of intermediate steps or conventional appearance consistency. It is the most upvoted paper on Hugging Face today with 151 votes.
  2. 2 OmniScientist: An Omni-Modal Omni-Discipline AI Scientist 83 UPVOTES · BOBO LI ET AL. · NATIONAL UNIVERSITY OF SINGAPORE · ARXIV 2608.13558 OmniScientist is an end-to-end AI scientist that runs multidisciplinary research across modalities instead of reasoning only over text, code, labels or precomputed summaries. The authors argue that existing systems leave out the spatial, temporal, cross-channel and procedural relations that scientific discovery actually turns on, so workflow coverage alone is not enough.
  3. 3 SPADE: Self-Play in Adaptive Synthetic Executable Environments 40 UPVOTES · BO LIU ET AL. · SPADE-RL · ARXIV 2608.19197 SPADE is a self-play reinforcement learning framework in which one language model takes two roles: an Environment Designer that writes long-horizon training environments as executable code behind a Gym-style reset and step interface, and a Reasoning Agent that learns inside them. The aim is a goal distribution that keeps expanding as the learner scales, which hand-curated or frozen environment pools cannot do.
  4. 4 Training Chemical Plausibility-Aware Large Language Models for Single-Step Retrosynthesis 29 UPVOTES · BOGDAN ZAGRIBELNYY ET AL. · INSILICO MEDICINE · ARXIV 2608.18940 Insilico Medicine trained C3LM, a chemistry model for single-step retrosynthesis, on a dataset of roughly 45.6 million verified reactions. The team proposes Top-K prompting to capture the intrinsically one-to-many nature of the problem, which single-answer benchmarks measure poorly, and pairs fine-tuning with plausibility and novelty rewards.
  5. 5 EDITBRIDGE: Towards Faithful and Efficient Ultra-High-Resolution Image Editing 21 UPVOTES · JIAYI SONG ET AL. · ARXIV 2608.18063 EDITBRIDGE targets image editing above 1K resolution, where quadratic attention cost and memory push most diffusion models into a two-stage workaround: edit small, then upscale. The authors say that pipeline hallucinates details contradicting the high-resolution source and leaves texture over-smoothed or over-sharpened, and propose editing at full resolution instead.

2026-08-19 · WEDNESDAY · 13:05 PDT

  1. 1 Demystifying Agent Skills: Why They Work-Until They Don't 102 UPVOTES · ZHIYUAN JIANG ET AL. · UNIVERSITY OF CALIFORNIA AT SAN DIEGO · ARXIV 2608.14036 A UC San Diego team ran controlled experiments to isolate when skills actually help LLM agents and when they fail. The study varies representation, outcome annotation, retrieval difficulty and cross-framework robustness across several benchmarks, agent harnesses and models. It argues aggregate task-success numbers hide where skills break down.
  2. 2 Agentic ESOpt: Fine-Tuning Long-Horizon LLM Agents with Minimal GPU Requirements 90 UPVOTES · ZHI ZHENG ET AL. · NATIONAL UNIVERSITY OF SINGAPORE · ARXIV 2608.17310 Researchers at the National University of Singapore argue for evolution strategies over reinforcement learning when fine-tuning long-horizon LLM agents. They say ES avoids the heavyweight backpropagation stack that keeps RL from scaling to larger models, and sidesteps the credit-assignment problem that branching, sparsely rewarded trajectories create.
  3. 3 Embodied-Navigator: Point, Think, Memorize, and Align for Efficient Navigation 43 UPVOTES · HONGYAN FENG ET AL. · ZJU-OMNIAI · ARXIV 2608.17512 A ZJU-OmniAI team introduces TAMP-Nav, which recasts embodied navigation as 2D visual prompting so a vision-language model only has to select pixels. The formulation avoids the unnatural action spaces that clash with a VLM's 2D pre-training priors, and pairs it with more flexible reasoning schedules and memory management.
  4. 4 AVA-Encoder: Towards Agent-Native Video Representation Learning 35 UPVOTES · CHUYUE LI ET AL. · QWEN BUSINESS UNIT · ARXIV 2608.12313 The Qwen Business Unit proposes AVA-Encoder, which converts a video into a knowledge-graph representation and then reconstructs the video from it. The agentic auto-encoding setup aims to give creative agents a structure they can reason over and manipulate, rather than an opaque latent, so they can learn from high-quality human films.
  5. 5 An Empirical Study of Training Pixel-Space Text-to-Image Diffusion Models 30 UPVOTES · DENGYANG JIANG ET AL. · TONGYI-MAI · ARXIV 2608.16887 Tongyi-MAI reports a large-scale study of text-to-image diffusion trained directly in pixel space rather than in a latent space. The authors find that direct pixel-space pre-training converges substantially more slowly at scale, and propose a latent-to-pixel strategy that acquires the representation in latent space first.

2026-08-18 · TUESDAY · 13:07 PDT

  1. 1 HarnessEval-W: Agentifying the Evaluation of Visual Worlds 109 UPVOTES · WEILIANG CHEN ET AL. · MIRROS · ARXIV 2608.16859 HarnessEval-W is an evaluation pipeline that judges world model rollouts with an agent harness instead of brute-force metrics. It produces a reasoning chain justifying each score, so violations of physics, causality and world state can be examined and verified rather than reduced to a scalar. It is the day's most upvoted paper at 109 votes.
  2. 2 StateM: Reaching 95.3% Raw Accuracy, or a $15 Frontier Run, on Terminal-Bench 2.1 via Harness Scaling 51 UPVOTES · ZIHENG QIN ET AL. · ARXIV 2608.15089 StateM is an agent-native runtime that organizes long-horizon execution around durable states, phase-local context, checked transitions and recoverable runbooks, leaving model weights untouched. The authors report 95.3 percent raw accuracy on Terminal-Bench 2.1 and a frontier run costing 15 dollars, framing it as a case for scaling the harness rather than the model.
  3. 3 Large Discovery Models: Empirically-grounded Model-Based Open-Ended Search 48 UPVOTES · ZHONGWEI YU ET AL. · YANGTZE-AILAB · ARXIV 2608.15669 The Large Discovery Model is a recurrent architecture for open-ended search over structured hypothesis spaces such as molecules, protein sequences and programs, where each candidate is expensive to evaluate. The authors argue that language model likelihoods and self-assessments are unreliable proxies for the objective and for calibrated uncertainty, and ground the search empirically instead.
  4. 4 Learn What's Left, Not What's Mastered: Saturation Aware Advantage Reweighting for Multi-Reward Policy Optimization 43 UPVOTES · YIXUAN WANG ET AL. · UNIVERSITY OF CALIFORNIA AT SAN DIEGO · ARXIV 2608.16072 The paper targets multi-reward reinforcement learning for post-training reasoners, where reward vectors are usually collapsed into a fixed weighted sum before group-wise standardization. The authors show this hands identical advantages to rollouts with different reward profiles and keeps spending gradient on objectives that are already saturated.
  5. 5 MOSS-VL Technical Report 38 UPVOTES · PENGYU WANG ET AL. · OPENMOSS · ARXIV 2608.15045 OpenMOSS released MOSS-VL, an open vision-language model family designed to perceive while it speaks. The language decoder attends to vision only through gated cross-attention, so incoming frames can be handled during generation, and a synthesized interaction corpus supervises when to speak, stay silent or revise. Real-time training sits in one light final stage over an offline foundation.

2026-08-17 · MONDAY · 13:06 PDT

  1. 1 Self-Supervised Visual On-Policy Distillation 144 UPVOTES · YIJIANG LI ET AL. · UNIVERSITY OF CALIFORNIA AT SAN DIEGO · ARXIV 2608.14144 The paper asks where the teacher-student asymmetry in visual on-policy distillation can come from when no privileged supervision is available. Instead of handing the teacher extra information such as reference answers or ground-truth regions of interest, the authors subtract information from the student, which they report yields the same effective learning signal for free.
  2. 2 Apodex Discovery: Reality Benchmarks and Environments for Evaluating and Building Discoverative Artificial Intelligence 22 UPVOTES · BRIAN WANG ET AL. · APODEX · ARXIV 2608.11341 Apodex Discovery is a framework of benchmarks and environments for AI aimed at open-ended discovery rather than pre-specified tasks, built around what the 29 authors call a heavy-duty solver. Their argument is that frontier models already handle hard problems once the goal, tools and success criteria are executable, and the missing step is turning real-world challenges into that form.
  3. 3 Marionette: Predicting World States, Rendering Geometry, Painting Appearance 22 UPVOTES · ZIAN MENG ET AL. · ARXIV 2608.14530 Marionette splits an interactive game world model into three jobs: predicting the evolving world state, computing geometry with a fixed zero-parameter renderer, and leaving the neural network to synthesize appearance. The design targets long-horizon drift, where models that autoregress pixels or latents accumulate errors in pose, geometry and occlusion until control breaks down.
  4. 4 DFM Mimir v1: An Open HRM Delivering Frontier Performance at 1B Parameters Using Only Permissible Post-Training Data 21 UPVOTES · PETER SCHNEIDER-KAMP ET AL. · UNIVERSITY OF SOUTHERN DENMARK (SDU) · ARXIV 2608.13517 Mimir v1 is a 1-billion-parameter model built on the Hierarchical Reasoning Model architecture and trained from scratch on a mixture of 161 permissibly licensed datasets. It outperforms the original HRM-Text 1B and sets a new state of the art for Danish, which the authors offer as evidence that ethically sourced data need not forfeit competitive results at this scale.
  5. 5 CPI-Bench: A Comprehensive, Practical and Intelligent Benchmark for Real-World Image Editing 12 UPVOTES · QINYE ZHOU ET AL. · ALIBABA · ARXIV 2608.14546 CPI-Bench is an image editing benchmark built for deployment-like conditions, covering multi-image edits, demanding reasoning instructions and practical settings. The 20 authors argue that existing benchmarks stay confined to simple single-image tasks, which leaves them unable to separate the performance of today's editing models.

2026-08-15 · SATURDAY · 13:05 PDT

  1. 1 LiveAnimate: Stable Long-Form Streaming Human Animation in Real-Time 12 UPVOTES · YUXUAN ZHANG ET AL. · QWEN BUSINESS UNIT · ARXIV 2608.11745 LiveAnimate turns a single reference image and a driving pose stream into human animation in real time, aimed at live streaming, telepresence and virtual avatars where diffusion systems need minutes per clip. It is built on a 14B-parameter video diffusion transformer with two-stage training, which the authors call the first system to pair real-time streaming with stable long-form generation at that scale.
  2. 2 Full-bandwidth transformer 10 UPVOTES · XI WANG ET AL. · MICROSOFT RESEARCH · ARXIV 2608.08888 Microsoft Research proposes widening the vertical feedback channel in autoregressive transformers, where today only the sampled token returns to the bottom of the stack and the top-layer hidden state is discarded. The full-bandwidth transformer fuses that previous hidden state with the sampled token embedding through a gated linear unit at each decoding step.
  3. 3 An AI4AI Framework for Visual Token Pruning 7 UPVOTES · ZHEN LIU ET AL. · ARXIV 2608.07193 Visual token pruning cuts the inference cost of multimodal LLMs, but the methods in use rely on hand-tuned heuristics and expert trial and error. This paper asks whether large language models can design the pruning algorithms themselves, and proposes an AI4AI framework that searches the design space as pruning objectives, budgets and architectures multiply.
  4. 4 H2R-Bench: Benchmarking Human-to-Robot Manipulation Video Generation in World Models 7 UPVOTES · DINGYI RONG ET AL. · SHANGHAI JIAO TONG UNIVERSITY · ARXIV 2608.13049 Robot demonstration data is expensive to collect, while egocentric human manipulation video is abundant but hard to transfer across embodiments. H2R-Bench, from Shanghai Jiao Tong University, evaluates how well video world models synthesize robot-centric manipulation clips from human footage, a cross-embodiment capability the authors say remains largely untested.
  5. 5 Thought-Level Beam Search for Reasoning 6 UPVOTES · LIJIE YANG ET AL. · PRINCETON UNIVERSITY · ARXIV 2608.08020 Princeton researchers frame test-time reasoning as a compute allocation problem over partial trajectories rather than a question of how much compute to spend. They argue that parallel sampling treats traces independently and creates memory bottlenecks, while subtractive pruning starves promising branches, and propose beam search over whole thoughts instead of tokens.

2026-08-14 · FRIDAY · 13:05 PDT

  1. 1 Alaya-EVOKE: From Linear-Scaling Supervision to Endless World 81 UPVOTES · YUANYANG YIN ET AL. · ARXIV 2608.13546 Evoke is an interactive world model that externalizes persistent world state instead of holding history in the denoiser context or key-value cache, whose cost grows with session length. Scene geometry is maintained outside the model and the teacher is redesigned for long-horizon interactive generation. The paper drew 81 upvotes, the most of the day.
  2. 2 DarwinX: Evolving Agent Harnesses Through Natural Selection 54 UPVOTES · YIFAN ZHANG ET AL. · SALESFORCE AI RESEARCH · ARXIV 2608.07545 DarwinX treats agent self-improvement as selection over a population of harnesses - prompts, tools, skills and control flow - with the model weights frozen. A preserve-and-extend contract admits only variants that extend coverage without regressing other tasks, while an archive keeps alternative lineages available for recombination.
  3. 3 How Can Rhetoric Reward-Hack AI Reviewers? Dissecting Rhetorical Sensitivity in AI-Based Peer Review 38 UPVOTES · MING LI ET AL. · UNIVERSITY OF MARYLAND · ARXIV 2608.08975 The authors rewrote 120 anonymized ICLR 2026 submissions along six rhetorical dimensions while preserving the reported scientific content, producing a controlled corpus of 4,200 manuscripts. Five LLM reviewers then scored the results under standard and strict protocols, measuring how far presentation alone shifts AI review judgments.
  4. 4 Spatial Memory Agent: Experience-Grounded Procedure Memory for Spatial Intelligence 27 UPVOTES · HAOKAI ZHANG ET AL. · ZHEJIANG UNIVERSITY · ARXIV 2608.12743 The paper asks whether a frozen vision-language model can improve its spatial reasoning through accumulated experience alone. It proposes procedure memory grounded in past attempts as a third route, complementary to post-training methods such as fine-tuning and to agents that call external depth and 3D reconstruction tools.
  5. 5 Massive Activations in Hybrid Linear Attention Large Language Models: Pre-Attention Spikes and Inter-Spike Plateaus 16 UPVOTES · ZUNHAI SU ET AL. · STARTLUX · ARXIV 2608.12149 This is the first systematic study of massive activations in layer-interleaved hybrid linear attention models, and it identifies two recurring shapes: spikes that appear immediately before full attention layers, and plateaus that persist across the intervening linear attention layers. The organization recurred across five linear attention architectures.

2026-08-13 · THURSDAY · 13:05 PDT

  1. 1 Spark-to-Paper: End-to-End Research Paper Generation as a Composable Skill 176 UPVOTES · ZHUOYANG QIAN ET AL. · ARXIV 2608.11924 Spark-to-Paper turns a research idea into a complete paper through thirteen composable skills running inside an existing coding assistant, with no separate agent platform or orchestration service. The system retrieves literature, designs and runs experiments, revises claims against evidence and produces publication-ready figures, separating model judgment from deterministic operations.
  2. 2 Mechanist: AI as a Scientific Instrument for Discovering the Mechanisms of Intelligence 73 UPVOTES · MENGRU WANG ET AL. · ZJUNLP · ARXIV 2608.12036 Mechanist is an agentic system that uses AI itself as a scientific instrument to autonomously discover the mechanisms behind model capabilities. The 19-author team argues that mechanistic exploration remains largely manual even as AI development accelerates and automates, widening the gap between what models can do and what researchers can understand or control.
  3. 3 SkillZip: Contract-Preserving Graph Compression for Scalable Agent Skill Libraries 72 UPVOTES · XINGYU TAN ET AL. · ARXIV 2608.05604 SkillZip compresses the reusable skill packages that LLM agents load at inference time, so the smallest sufficient executable context fits a limited budget. The authors trace current failures to a unit mismatch, where skills are retrieved as whole packages but reused below that level, and compress as a graph that preserves procedural contracts and keeps routines executable.
  4. 4 Articulated Object Reconstruction from Rest-State Observation 42 UPVOTES · DAEUN LEE ET AL. · SEOUL NATIONAL UNIVERSITY · ARXIV 2607.27749 The paper reconstructs articulated objects from a single closed configuration, recovering both 3D geometry and the kinematic structure that governs how parts move. Existing methods need observable motion across multiple articulation states, so the Seoul National University team leans on geometry, semantics and motion priors, using an explicit mesh for cross-model verification.
  5. 5 AdvFD: Boosting Visual Generation via Adversarial Frechet Distance Loss 24 UPVOTES · MINGJU GAO ET AL. · KOLORS TEAM, KUAISHOU TECHNOLOGY · ARXIV 2608.11205 AdvFD attacks Frechet hacking, where a generator keeps improving its target Frechet score while visual quality and alignment in other feature spaces stagnate or degrade. The Kuaishou Kolors team blames the static pretrained feature spaces used by existing Frechet losses, which give a fixed and incomplete view of the gap between real and generated distributions.