xClean Tools

AI Papers — Daily Top 5

UPDATED 2026-09-05 13:05 PDT

2026-08-28 · FRIDAY · 13:05 PDT

  1. 1 Agentic Game Development as a Verifiable Trajectory Data Engine for Scaling World Models 117 UPVOTES · PENGFEI ZHOU ET AL. · NATIONAL UNIVERSITY OF SINGAPORE · ARXIV 2608.25518 The authors argue that scaling world models on more crawled video and more compute is inefficient without grounded reward signals, and propose a recursive data engine built on agentic game development instead. Because games are executable, the trajectories they produce can be scored the way compilers score code agents, rather than through fuzzy proxies such as CLIP scores.
  2. 2 UrbanGround: From Local Perception to Spatial Agency in a Real-Scale City 68 UPVOTES · TIANJIE JU ET AL. · SHANGHAI JIAO TONG UNIVERSITY · ARXIV 2608.27456 UrbanGround tests whether multimodal LLM agents can turn street-level perception into reliable action once they begin to move through a city. The sandbox is a physically constrained replica of Hong Kong built from territory-wide 3D geospatial data, supporting closed-loop first-person interaction and an interactive map for navigation.
  3. 3 TTPO: Test-Time Policy Optimization 62 UPVOTES · AOZHE WANG ET AL. · ARXIV 2608.27448 Post-training methods such as reinforcement learning and on-policy self-distillation depend on ground-truth labels, which rules out test-time training, and majority-vote pseudo-labels are fragile because one wrong vote corrupts the teacher. The authors find the failure is asymmetric: rollouts that disagree with the pseudo-label are usually wrong whether or not the vote itself is, and TTPO builds on that.
  4. 4 Self-OPD: On-Policy Distillation for Flow Matching Models without Teacher 55 UPVOTES · SHIYI ZHANG ET AL. · ARXIV 2608.26872 On-policy distillation gives dense supervision but normally requires a specialized teacher trained for every new objective, and the gap between teacher and student distributions compounds errors along the generation trajectory. Self-OPD is a teacher-free framework that keeps the dense signal for flow matching models while removing both costs.
  5. 5 What Makes Good Agentic Data? An ACE Lens on Data Generation for LLM Agents 55 UPVOTES · XINGSHAN ZENG ET AL. · ARXIV 2608.27260 Work on generating interaction data for LLM agents is organized by domain and evaluated inconsistently, which obscures the mechanisms the methods share and blurs candidate construction into verification and selection. This paper proposes a two-level framework for the field, treating consistency among environments, tasks, interactions and success signals as what generation must preserve.

2026-08-27 · THURSDAY · 13:05 PDT

  1. 1 VGI-BENCH: Probing Visual Intelligence in Video Generation Models 139 UPVOTES · XUAN HE ET AL. · UNIVERSITY OF ILLINOIS AT URBANA-CHAMPAIGN · ARXIV 2608.19583 A 22-author team led from the University of Illinois at Urbana-Champaign introduced VGI-bench, a benchmark for the zero-shot visual reasoning that video generation models can show through their generated frames. It holds 27 tasks and 810 instances under a two-level taxonomy of task domains and skill tags. Tasks demand a valid evolving process, not just a plausible final state.
  2. 2 JIT-Agent: Scaling Harness Intelligence via Just-in-Time Harness Evolution 47 UPVOTES · GUIBIN ZHANG ET AL. · NATIONAL UNIVERSITY OF SINGAPORE · ARXIV 2608.25593 Researchers at the National University of Singapore presented JIT-Agent, a model trained to synthesize an agent harness on the fly for any off-the-shelf agentic LLM. It formalizes the harness - memory, planning, action protocol, tool and skill orchestration - as a composable, machine-generatable artifact. The premise is that harness design, still manual, can dominate the contribution of the underlying model.
  3. 3 CyberFactory: Scaling Cyber Security Capabilities with Instances from the Wild 30 UPVOTES · JIAN YANG ET AL. · IQUEST · ARXIV 2608.23181 A team at IQuest released CyberFactory, a unified open-source framework for building cybersecurity capability into large language models. The paper argues open efforts trail closed models here because frontier open-weight releases ship no reproducible security training recipe, while existing open work covers isolated tasks and lacks scalable agentic data.
  4. 4 D^3-MOPD: Adaptive Dynamic Domain ScheDuling for Efficient Multi-Teacher Distillation 22 UPVOTES · ZECHEN SUN ET AL. · ARXIV 2608.24987 A ten-author team proposed D^3-MOPD, which adapts the per-domain data mixture during multi-teacher on-policy distillation instead of fixing it before training. Their motivation is that domains converge at very different rates, so a fixed mixture spends compute on domains that plateaued early while undertraining the slower ones.
  5. 5 Agent-G^2: Gaussian Guidance for Agentic Reinforcement Learning 20 UPVOTES · ZIXUAN WANG ET AL. · ZHEJIANG UNIVERSITY · ARXIV 2608.23318 Zhejiang University researchers proposed Agent-G^2 for hint-based reinforcement learning, in which a prefix of an expert trajectory is retained before each rollout so the policy explores from a state closer to success. Existing methods treat that guidance depth as one deterministic number; the authors find useful guidance instead occupies a band of depths.

2026-08-26 · WEDNESDAY · 13:05 PDT

  1. 1 GigaBrain-0.7: Scaling Embodied Foundation Models to Emergent Capabilities with a Three-System Architecture 89 UPVOTES · GIGABRAIN TEAM ET AL. · GIGAAI · ARXIV 2608.15875 GigaAI presents GigaBrain-0.7, an embodied foundation model organized around a three-system architecture. The paper asks whether current vision-language-action systems still gain from better architectural design and from far larger, more heterogeneous training data, and reports improved generalization across different robot embodiments and tasks.
  2. 2 AutoSaddler: Automatic Harness Optimization with Durable Updates from Agent Execution Traces 49 UPVOTES · SUNGHO PARK ET AL. · MICROSOFT · ARXIV 2608.23041 Microsoft proposes AutoSaddler, which recasts agent harness design as an offline learning problem rather than manual search over prompts, tool configurations and control logic. It reads failure signals from execution traces in mini-batches and iteratively updates the harness. The target is long-horizon tasks, where small local failures compound into overall failure.
  3. 3 On-Policy Self-Distillation in Diffusion Models 47 UPVOTES · WEI ZHOU ET AL. · BYTEDANCE SEED · ARXIV 2608.24646 ByteDance Seed introduces DiffusionOPSD, which converts image-level reward guidance into explicit targets for a diffusion model's clean-output predictions. A frozen behavior policy supplies the trajectories and anchors, and reward gradients build bounded positive and negative targets around each anchor. The point is to specify how an intermediate denoising prediction should change, which endpoint rewards do not.
  4. 4 SecOPD: Mitigating Adaptive Prompt Injections by On-Policy Distillation 36 UPVOTES · YIBO PENG ET AL. · ARXIV 2608.21500 The authors argue that defensive finetuning still fails against adaptive prompt injection because DPO and GRPO grade an entire output with one sequence-level signal, treating every token alike. SecOPD uses on-policy distillation to supply token-level feedback instead. Prompt injection, in which text hidden in fetched pages, files or email redirects an agent, is listed as the top threat to AI agents.
  5. 5 The Mask Is Not the Model: Auditing Prefix Invariance in Attention, State-Space, and Hybrid Sequence Models 29 UPVOTES · TAEBONG KIM ET AL. · VIDRAFT · ARXIV 2608.22876 The paper formalizes prefix invariance - a representation at position t must not depend on later inputs - and gives a two-forward-pass audit, with no training or gradients, that localizes where causality breaks. Mask inspection is not enough: leaks can travel through scans or normalization. Across 192 injected-fault trials it caught none, while the audit found all 192 plus a defect in Zamba2 and Nemotron-H.

2026-08-25 · TUESDAY · 13:05 PDT

  1. 1 TLive-Omni: An Omni-Modal Understanding Model for E-Commerce Live Streaming 52 UPVOTES · YIBO HU ET AL. · TAOLIVE AIGC · ARXIV 2608.20958 TLive-Omni is an omni-modal model built for e-commerce live streams, where product facts are scattered across speech, video frames, product images, overlaid text and viewer questions. It maps image, video, audio and text into a single representation space, and adds Per-vGrid, a timestamped token layout that pairs each video grid with its matching audio inside explicit boundary tokens.
  2. 2 InfinityEdit: Infinite Video Editing with a Lightweight Edit-Ignition Adapter 35 UPVOTES · YUNZE TONG ET AL. · ARXIV 2608.20910 Instruction-based video editing usually assumes an in-place edit, aligning the output frame by frame with a fixed source clip. InfinityEdit takes on what the authors call infinite video editing, where an instruction given on a preceding segment must carry into frames that arrive later, as when restyling a live game or extending a camera move on an ongoing shot.
  3. 3 MobilePA-Bench: Benchmarking Mobile Planner Agents on Complex Real-World Tasks 33 UPVOTES · YI ZHU ET AL. · TONGYI-MAI · ARXIV 2608.23035 MobilePA-Bench is an interactive, stateful, tool-centric benchmark for on-device LLM agents acting as phone copilots. The authors argue existing tests fall into two camps with matching blind spots: GUI benchmarks that measure only surface screen manipulation, and static function-calling sets that match APIs offline, away from real runtime constraints.
  4. 4 Prime Agent: A Self-Improving RLM Harness 32 UPVOTES · SETH KARTEN ET AL. · PRIME INTELLECT · ARXIV 2608.23552 Prime Intellect describes Prime Agent, an open-source harness for long-horizon evaluation and coding-agent workflows. A persistent IPython REPL implements the Recursive Language Model abstraction for programmatic context processing and test-time compute, while a Continual Harness carries histories, memories, skills, prompts and subagent specifications across trajectories.
  5. 5 Block3D: Efficient Text-to-3D Generation via Block-Wise Diffusion 25 UPVOTES · BOWEN CUI ET AL. · ZHEJIANG UNIVERSITY · ARXIV 2608.19567 Block3D generates 3D shapes from text using a block-wise diffusion framework that partitions the discrete shape representation instead of decoding tokens one at a time or refining the whole representation at every step. The authors note that autoregressive decoding cannot revise its own errors, while diffusion and flow models pay full-representation cost each iteration.

2026-08-24 · MONDAY · 13:04 PDT

  1. 1 ParaTempo: Efficient Parallel Reasoning via Temporal Confidence 26 UPVOTES · XUTENG ZHANG ET AL. · SHANGHAI JIAO TONG UNIVERSITY · ARXIV 2608.16425 ParaTempo is a training-free framework for asynchronous parallel reasoning in large reasoning models. The authors argue that existing controls for parallel branches - final-answer consensus, local token confidence, or isolated intermediate probes - are delayed, weakly tied to reasoning progress, or too noisy for branch-level decisions, and propose a temporal confidence signal instead.
  2. 2 Beyond Correctness: Benchmarking and Aligning Response Behaviors in Hybrid-Thinking MLLMs 16 UPVOTES · XINMING WANG ET AL. · TENCENT · ARXIV 2608.12781 Tencent researchers examine multimodal models that switch between deliberative thinking and low-latency non-thinking inference, arguing that both modes should meet the same user-facing standard. The work evaluates response-pattern failures alongside task accuracy and tests whether the two interfaces preserve acceptable final-response behavior.
  3. 3 EviRank: Structured Relevance Evidence for Multimodal Image Re-ranking 10 UPVOTES · ENJUN DU ET AL. · ARXIV 2608.20886 EviRank recasts multimodal image re-ranking as a semantic constraint satisfaction problem, parsing text-only, image-only or composed queries into structured relevance evidence. The authors argue existing re-rankers either compress multifaceted relevance into an opaque embedding or lean on free-form chain-of-thought that drops or hallucinates fine-grained constraints.
  4. 4 UniSpace: Unified Visual Representation and Scalable Multimodal Modeling 8 UPVOTES · JINBO YAN ET AL. · LONGCAT · ARXIV 2608.08676 UniSpace asks whether understanding, generation and editing can share one visual representation space built from a pretrained semantic vision transformer. The authors note that final tokens from semantic encoders discard fine detail and hurt pixel reconstruction, and show that frozen ViT blocks are not inherently unable to preserve it.
  5. 5 Partition the Support, Reconstruct the Residual: Training-Free Sparse Attention for Video Generation and World Models 6 UPVOTES · PARDIS TAGHAVI ET AL. · TEXAS A&M UNIVERSITY · ARXIV 2608.18484 The paper introduces SparsePR, a training-free block-sparse attention method for video transformers that pairs response-coupled partitioning with probe-fitted residual reconstruction. The authors argue that row-wise attention concentration alone does not specify a usable sparse operator, because queries sharing a block route may have poorly overlapping supports.

2026-08-22 · SATURDAY · 13:04 PDT

  1. 1 Thinking in a Low-Resource Language: What SFT Builds, What RL Fixes, What Accuracy Cannot See 13 UPVOTES · AYOUB KIROUANE ET AL. · KIEFER · ARXIV 2608.17744 Three frontier mixture-of-experts models with 3.6 to 4.0B active parameters were fine-tuned to reason in Greek, and accuracy barely moved. Changing only the random seed shifted scores by 7.7 points, more than any data or recipe effect measured, making the benchmark itself noise at that scale. What did change was the language of thought: base models produced Greek reasoning in 0 of 1,000 traces.
  2. 2 Hierarchical Self-Improvement: A Framework for Task-Specific Evolvable Agent Harnesses 10 UPVOTES · TAILIN ZHOU ET AL. · HKUST · ARXIV 2608.08466 LLM agents are usually improved by editing prompts, tools or workflows, while the harness that executes around the model stays fixed after deployment. This work makes the harness task-specific and continuously evolvable: each task family keeps its own harness, hot-swapped across iterations through a fixed task-injection seam and rewritten from environment feedback, with a single frozen model throughout.
  3. 3 EXIMO: VLM Guided Exploration of VLA Policies 10 UPVOTES · BHAVYA SUKHIJA ET AL. · DEEPMIND · ARXIV 2608.19891 DeepMind takes on how to finetune robot policies for new tasks on the fly. State-of-the-art manipulation rests on behavior cloning of billion-parameter vision-language-action models trained on huge teleoperation datasets, and adapting them is still open: more teleoperation costs hundreds of hours of human labor, while reinforcement learning is sample-hungry. EXIMO has a vision-language model guide exploration.
  4. 4 Inject, Align, Recover: Staged Post-Training for Retrieval-Free Document Knowledge Internalization 10 UPVOTES · QIAN KOU ET AL. · BEIJING ACADEMY OF ARTIFICIAL INTELLIGENCE · ARXIV 2608.20281 Language models often fail on questions about a bounded document collection when the sources are not retrieved at inference time. IAR is a three-stage post-training framework that separates the problem: Inject turns source documents into continuation data, Align tunes question-answering behavior, and Recover restores general ability. The authors present it as an alternative to conventional continued pretraining.
  5. 5 NARU: A Benchmark for NARrative Evolution and Cultural Nuance Understanding in Japanese Extreme Long Video 8 UPVOTES · YUHENG HUANG ET AL. · THE UNIVERSITY OF TOKYO · ARXIV 2608.13210 NARU is a benchmark for tracking an evolving narrative and reading implicit social meaning in Japanese long-form video. It holds 1,481 questions grounded in 155 videos totaling 146.8 hours, spanning four narrative and five cultural dimensions. Existing benchmarks rarely test those two capabilities together, especially in high-context non-English media.

2026-08-21 · FRIDAY · 13:08 PDT

  1. 1 FACET: Preserving Source Intent and Executable State in Terminal Task Synthesis 102 UPVOTES · KOU SHI ET AL. · UNIVERSITY OF SCIENCE AND TECHNOLOGY OF CHINA · ARXIV 2608.18580 Training terminal agents needs executable tasks at scale, yet each task couples an instruction, an initialized environment, a reference solution and a verifier, and inconsistent assumptions across those parts leave tasks unsolvable or wrongly graded. FACET synthesizes tasks while carrying the goals, dependencies and state transitions of the original sources through the pipeline.
  2. 2 SWE-bench Science: Can Coding Agents Resolve Engineering Tasks in Science? 51 UPVOTES · ZHIPENG XU ET AL. · OPENMOSS · ARXIV 2608.19799 OpenMOSS introduced SWE-bench Science, a repository-level benchmark of 119 tasks drawn from 98 GitHub repositories across 20 scientific domains. It is built to show why coding agents fail when repairing scientific software rather than only whether they succeed, on the argument that faulty scientific code can compromise the evidence behind published conclusions.
  3. 3 WithEveryone: Unified Planning and Identity Grounding for Group Image Generation 34 UPVOTES · HENGYUAN XU ET AL. · TENCENT HUNYUAN · ARXIV 2608.20336 Tencent Hunyuan's WithEveryone generates group images holding up to ten specified people without their identities blurring together. Each reference is injected as an addressed token, the model predicts a structured identity and layout plan, and that plan is rendered back as a visual condition for generation.
  4. 4 MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use 24 UPVOTES · MENGRU WANG ET AL. · ZJUNLP · ARXIV 2608.20202 MemTrapBench targets what its authors call memory-induced cognitive traps: even a faithfully recorded and relevant memory can distort a model's reasoning and hurt performance on the task at hand. Existing memory benchmarks mostly check whether information was extracted, stored and retrieved correctly, not how the retrieved text reshapes the answer.
  5. 5 SkillEvo: Self-Renewing Evolution Gradients from Multi-Turn Interaction Feedback 21 UPVOTES · QIANXI YAN ET AL. · TENCENT · ARXIV 2608.13120 Agent skills are usually hand-authored or produced in a single generation pass, so they never learn from the failures they cause. Tencent's SkillEvo argues that recent feedback loops stall because they score skills on single-turn question answering, which hides defects that only surface across several turns, and instead draws its evolution signal from multi-turn interaction.