xClean Tools

AI Papers — Daily Top 5

UPDATED 2026-11-03 01:12 PST

2026-09-12 · SATURDAY · 13:05 PDT

  1. 1 An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics 22 UPVOTES · IVAN MOSHKOV ET AL. · NVIDIA · ARXIV 2609.10712 NVIDIA researchers post-trained Nemotron 3 Ultra with supervised fine-tuning and reinforcement learning to write natural-language proofs for olympiad mathematics. The resulting test-time-compute pipeline runs entirely in natural language, with no formal prover, external tools or internet access, and combines the base model with two specialist checkpoints. The paper publishes the recipe behind its IMO gold claim.
  2. 2 FreeFlow: A Bias-free Hierarchical Transformer for Optical Flow Estimation 22 UPVOTES · VLADISLAV BARGATIN ET AL. · ARXIV 2609.11486 FreeFlow estimates optical flow with a hierarchical transformer that drops the correlation volumes, feature warping and iterative refinement the field normally relies on. In their place it uses a single feed-forward encoder-decoder mixing several attention variants, including window and shifted-window attention. The authors argue task-specific inductive biases limit expressivity and add pipeline cost.
  3. 3 MetroLLM-Bench: Evaluating Language Models as Transit Kiosk Runtimes 19 UPVOTES · REMCO HENDRIKS ET AL. · CONTINKER · ARXIV 2609.10016 MetroLLM-Bench tests language models as the policy layer behind a transit ticket kiosk, with 955 cases across six real metro systems ranging from 37 to 414 stations. Models must call structured tools and return a machine-renderable terminal state carrying an outcome, a per-ticket fare quote where applicable and a kiosk action. Scoring splits into 14 deterministic components and eight semantic-quality ones.
  4. 4 TempCloze: Can Video-LLMs Identify the Missing Middle? 18 UPVOTES · WENQI PEI ET AL. · THE UNIVERSITY OF HONG KONG · ARXIV 2609.01515 TempCloze asks video language models to pick the true missing middle of a clip from four candidates, given only its opening and closing segments. The benchmark holds 1,521 filtered videos, mainly long-take and egocentric footage, with distractors built from the same source to remove option-wording and language-prior shortcuts. It aims at visual temporal reasoning rather than reasoning mediated by text.
  5. 5 Studying Image Tokenizers as Visual Languages in Unified Multimodal Models 18 UPVOTES · SITING LI ET AL. · AMAZON FRONTIER AI & ROBOTICS · ARXIV 2609.09143 The paper treats image tokenizers as the visual language of unified multimodal models and studies them in a controlled pure-autoregressive testbed. The authors track task-specific validation losses for text, image, text-to-image and image-to-text prediction during multimodal continual pretraining, and relate them to downstream performance. The goal is to judge tokenizers by how visual tokens behave with text.

2026-09-11 · FRIDAY · 13:06 PDT

  1. 1 SpatialBlock: Enhancing Spatial Intelligence in LVLMs via Synthetic Block-Stacking Problem 57 UPVOTES · SOOHYUN RYU ET AL. · KAIST AI · ARXIV 2609.07064 KAIST AI researchers propose SpatialBlock, which trains spatial reasoning in vision-language models on synthetic block-stacking problems instead of annotated real scenes. The paradigm is drawn from human cognitive development, and sidesteps the cost and label noise of real-scene spatial QA datasets that depend on external perception modules. It drew 57 upvotes, the most of any paper today.
  2. 2 WearableQA: A Benchmark for Health Reasoning over Real-World Wearable Data 31 UPVOTES · JI SOO LEE ET AL. · META · ARXIV 2609.05405 Meta introduced WearableQA, a benchmark of 4,084 ten-option questions built from the wearable time series, blood biomarkers and demographics of 200 real users, each with up to 500 days of daily measurements. It keeps the device noise and individual variation of real records rather than smoothing them, testing whether models can reason over a longitudinal health history. The paper drew 31 upvotes.
  3. 3 SWE-Bench Pro Verified: A Reliable Benchmark for Software Engineering Agents 21 UPVOTES · PUJUN ZHENG ET AL. · SHANGHAI AI LABORATORY · ARXIV 2609.08149 Shanghai AI Laboratory released SWE-Bench Pro Verified, a cleaned version of the SWE-Bench Pro agent benchmark. The authors report the original is undermined by reward hacking, where gold solutions or hidden evaluation information leak to the agent, and by task quality problems such as misleading problem statements and improperly scoped tests. Both inflate scores and obscure real coding ability.
  4. 4 Mi-Ripple: Restoring Images Degraded by Iterative AI Editing 20 UPVOTES · JIAYIN CHEN ET AL. · MIYANG TECHNOLOGY (SHANGHAI) CO., LTD. · ARXIV 2609.11317 Mi-Ripple targets the grid-like and granular texture, known as digital ripple, that builds up when an image is edited repeatedly by reference-conditioned models. The workflow first separates periodic lattice artifacts from texture entangled with content, then applies spectral notching, structure-aware smoothing or regeneration from a cleaned reference depending on the diagnosis. The paper drew 20 upvotes.
  5. 5 PARSER: Read in Parallel, Reason in Depth for Long-Context LLM Agents 17 UPVOTES · KUN LI ET AL. · ARXIV 2609.06702 PARSER decouples reading from reasoning in long-context agents. A bank of lightweight subagents, each bound to one chunk, reads an entire document in parallel while a lead agent runs iterative scatter-gather rounds, broadcasting a query and aggregating the returns. The design targets two weaknesses of sequential memory agents: sensitivity to where evidence sits, and latency that grows with document length.

2026-09-10 · THURSDAY · 13:05 PDT

  1. 1 Show-Harness: Just a VLM Agent Can Play Robots 96 UPVOTES · YANZHE CHEN ET AL. · SHOW LAB · ARXIV 2609.10522 Show-Harness is an embodied harness that lets vision-language models control robots through a compact semantic interface linking intent to action. The model reasons over discrete semantic action units, while embodiment-specific interpreters deterministically ground those units into local robot actions. The design keeps the VLM directly responsible for fine-grained physical decisions.
  2. 2 AgentGrad: Intervention-guided Prompt Optimization for Multi Agent Systems 64 UPVOTES · JAEWON CHU ET AL. · ARXIV 2609.08572 AgentGrad addresses prompt optimization for LLM-based multi-agent systems, where textual gradient methods update each agent's prompt from natural-language feedback. The authors identify limitations at two stages, gradient extraction and gradient aggregation, noting that prior work picks a target prompt without checking whether changing it actually resolves the failure.
  3. 3 Programmable World Model 62 UPVOTES · ZHENG-HUI HUANG ET AL. · ALAYA LAB · ARXIV 2609.10540 Programmable World Model separates world-state evolution from visual observation generation so that video world models can hold persistent state across long interactions. An agent translates natural-language instructions into executable programs that specify entity states and state-transition rules, and a lightweight engine runs those programs to update the world before anything is rendered.
  4. 4 SceneMosaic: Efficient and Diverse Simulation-Ready Scene Generation via Hybrid Agentic Layout Evolution 41 UPVOTES · XINGJIAN RAN ET AL. · THE UNIVERSITY OF HONG KONG · ARXIV 2609.05594 SceneMosaic generates diverse, simulation-ready indoor 3D scenes through what the authors call hybrid agentic layout evolution. It aims to combine the fidelity of VLM-driven agentic pipelines, which require costly iterative object placement and refinement, with the speed of parametric image-to-3D models, which are efficient but often yield imprecise or physically invalid scenes.
  5. 5 BeaconKV: Key-Value Cache Compression Guided by Beacon Queries for Efficient Large Reasoning Model Inference 32 UPVOTES · JANGHYEON KIM ET AL. · ARXIV 2609.04971 BeaconKV compresses the key-value cache of large reasoning models, whose caches grow linearly with the long chain-of-thought traces they generate and often exceed GPU capacity. The authors show that existing methods, which estimate future token importance from recent queries, break down in long-horizon reasoning at decoding steps that revisit earlier thoughts.

2026-09-09 · WEDNESDAY · 13:06 PDT

  1. 1 AuK Technical Report: An Open-Source Foundational Model for Speech Generation and Editing 163 UPVOTES · ZIYANG MA ET AL. · TENCENT HUNYUAN · ARXIV 2609.08936 Tencent Hunyuan released AuK, an open-source foundational model that unifies speech generation and editing behind natural-language instructions and audio context. The team reports about 3.03 billion instruction-audio instances and 1.95 million hours of effective supervision across five task families, from content editing to enhancement, separation and acoustic editing. It is the day's most upvoted paper at 163 votes.
  2. 2 OpenWAM: An Open, Modular Exploration Towards Systematic World-Action Model Pretraining 48 UPVOTES · YURAN WANG ET AL. · OPENWAM · ARXIV 2609.07398 OpenWAM is an open research stack for pretraining world-action models, which turn video-generative priors into executable robot control. The authors argue existing systems are monolithic, with backbone, visual representation, architecture and training data tightly coupled, and factorize that design space into composable modules under unified training and inference. The aim is to show which design choices matter.
  3. 3 Marigold V2: Revisiting Diffusion Transformers for Monocular Depth Estimation 44 UPVOTES · IGOR PAVLOVIC ET AL. · HUAWEI BAYER LAB · ARXIV 2609.08084 Marigold V2 revisits the recipe for turning diffusion transformer image models into monocular depth estimators, this time targeting single-step inference. The authors note that despite the field's maturity, current models still fail to generalize to out-of-distribution inputs and produce blurry depth maps. Depth estimation feeds scene reconstruction, computational photography and robotics.
  4. 4 Miles v0.1: Production-Level Post-Training 44 UPVOTES · RADIXARK ET AL. · RADIXARK · ARXIV 2609.08368 RadixArk published Miles v0.1, a full-stack system for frontier post-training built on the design of slime. Rollout engines run on SGLang and the trainer offers a choice of two backends, one of them NVIDIA Megatron-LM, with every stage of the reinforcement-learning loop built to be verified, clean and customizable. The report walks the pipeline end to end.
  5. 5 Mask Forcing: Improving Autoregressive Video Diffusion Distillation via Dual-Noise Masking Rollout 43 UPVOTES · ZHUORAN ZHAO ET AL. · ARXIV 2609.09123 Mask Forcing targets the quality loss when bidirectional video diffusion models are distilled into causal autoregressive students for real-time generation. The authors trace the over-saturated, over-smoothed output to the mode-seeking behavior of the reverse KL objective in distribution matching distillation, which can collapse the student onto a few of the teacher's modes. The fix is a dual-noise masking rollout.

2026-09-08 · TUESDAY · 13:07 PDT

  1. 1 Dr. Claw: An AI Scientist Workspace for Vibe Research 123 UPVOTES · DINGJIE SONG ET AL. · ARXIV 2609.00365 The paper presents Dr. Claw, an open-source workspace that wraps command-line coding agents such as Claude Code and Gemini CLI in a human-in-the-loop workflow instead of adding another autonomous agent. Persistent state objects, a reusable skill library and multi-executor coordination link human decisions to agent execution. It is the day's most upvoted paper at 123.
  2. 2 Unlocking Lossless Speedups in LLMs via Discrete Diffusion 67 UPVOTES · SUBHAM SEKHAR SAHOO ET AL. · INSTITUTE OF FOUNDATION MODELS · ARXIV 2609.04010 The paper introduces diffusion-augmented LLMs, which keep an autoregressive model distribution but use diffusion to draw several tokens from it in parallel. Parameters are split into AR weights trained with standard next-token prediction and lightweight diffusion weights trained to emit multiple tokens at once. The aim is to break the sequential decoding bottleneck without changing the model's distribution.
  3. 3 FlowBalance: Verifier-Grounded Self-Improvement from On-Policy Reasoning Experience 58 UPVOTES · ZIXUN HUANG ET AL. · TENCENT HUNYUAN · ARXIV 2609.03241 FlowBalance is a verifier-grounded self-improvement method for reasoning models: it learns a normalized distribution over complete responses rather than leaning on sparse terminal verifiers alone. A frozen view of the same policy, given privileged context, produces token-level log-probability gains that are aggregated per trajectory. It targets dense same-model guidance that reinforces false confidence.
  4. 4 Iris: Climbing to the Search Frontier 53 UPVOTES · ZIYUAN LIU ET AL. · ALLSPARK RESEARCH · ARXIV 2609.04304 AllSpark Research presents Iris-mini and Iris-pro, two search agents trained at the 35B-A3B and 397B-A17B scales, with the data pipeline and recipe behind them. Tasks are reverse-constructed from a web corpus's hyperlink structure: multi-hop chains over an entity graph, with each non-answer entity rewritten as a descriptive reference so no clue yields to string matching.
  5. 5 Motion-Omni: End-to-End Joint Speech and Full-Body Motion for Spoken Dialogue 38 UPVOTES · CHENGQIAN MA ET AL. · PEKING UNIVERSITY · ARXIV 2609.04250 Motion-Omni is an end-to-end framework in which a spoken dialogue model natively outputs both speech and full-body motion, rather than cascading a motion model over already-generated audio. That cascade needs a second full inference pass and rules out joint optimization between what an avatar says and how it moves. The work comes from Peking University.

2026-09-07 · MONDAY · 13:03 PDT

  1. 1 Bilevel Coordinated Reflection: A Game-Theoretic Approach to Multi-Agent LLM Systems 91 UPVOTES · YIHANG CHEN ET AL. · ARXIV 2609.02750 The paper models orchestrator-worker interaction in multi-agent LLM systems as a bilevel coordination game. Under bounded coupling it treats the workers' local-update game as an approximate potential game whose equilibrium slack is controlled by decomposition quality, and analyzes textual reflection as stochastic movement over semantic memory states. It is the day's most upvoted paper at 91.
  2. 2 The Attention Triangle in Audio-Video Models 25 UPVOTES · SAGI POLACZEK ET AL. · TEL AVIV UNIVERSITY · ARXIV 2609.03586 Tel Aviv University researchers probe what they call the attention triangle in audio-video diffusion models: the three cross-attention edges linking text, audio and video. Routing along the audio-video edge turns out to be bidirectional, so audio can influence generated video and video can influence generated audio. The same mechanism that coordinates the modalities also leaks semantics between them.
  3. 3 Enoki: Efficient Multi-Level Hallucination Detection 16 UPVOTES · ELISEI RYKOV ET AL. · S-NLP · ARXIV 2609.00581 Enoki is an Open Information Extraction framework that detects LLM hallucinations at both the claim and span level. Existing detectors work at only one of the two, and bridging them is expensive: LLM-heavy pipelines need repeated decomposition and verification calls, while modular systems need claim-to-span alignment. Enoki instead extracts text-anchored relational facts and verifies those.
  4. 4 Beneath the Surface of Chains-of-Thought: A Mechanistic Interpretation of Reasoning Operations in LLMs 10 UPVOTES · SEOGYEONG JEONG ET AL. · NAVER AI LAB · ARXIV 2609.04753 NAVER AI Lab asks whether the functional operations of chain-of-thought reasoning, such as problem formulation, goal decomposition and deduction, are geometrically organized in a model's hidden representations. The operations prove separable in held-out representations, with separability peaking in middle layers, and the authors verify the structure is not explained by lexical cues.
  5. 5 MaxKernel: Agentic Kernel Generation for TPUs 9 UPVOTES · SHANGKUN WANG ET AL. · GOOGLE · ARXIV 2609.04523 Google presents MaxKernel, a multi-agent system that writes high-performance custom TPU kernels using real-time compiler feedback. It implements three paradigms: a human-in-the-loop agent for step-by-step design, an autonomous agent running a metric and trace driven optimization loop, and a graph-based agent. The target is work that normally demands deep hardware expertise.

2026-09-05 · SATURDAY · 13:05 PDT

  1. 1 RoboTok: An Internet-Scale Data Engine for Human Demonstration Retrieval and Dexterous Manipulation Learning 62 UPVOTES · HOWARD QIAN ET AL. · RICE ROBOTPI LAB · ARXIV 2609.03199 RoboTok is a data engine that takes a query video of a human manipulating an object and retrieves matching demonstrations from web video at internet scale, then trains dexterous robot policies on them. It learns a latent motion space from 3D hand trajectories in actor-centered reference frames, so retrieval transfers across people and scenes. The target is the long tail of tasks robot data collection cannot cover.
  2. 2 Scal3R: Learning Efficient Multi-Relative Pose Query for Scalable Online 3D Reconstruction 45 UPVOTES · CHIN-YANG LIN ET AL. · NATIONAL YANG MING CHIAO TUNG UNIVERSITY · ARXIV 2609.04201 Online 3D reconstruction degrades on long videos, and the authors trace the failure to regressing every pose against a fixed first-frame anchor, which pushes the model far outside its training distribution. Per-frame depth stays intact while the global pose head collapses, so Scal3R reframes the task as querying poses relative to multiple references. The change keeps reconstruction stable over much longer sequences.
  3. 3 Editable Visual Design 40 UPVOTES · JUNYAN YE ET AL. · TENCENT HUNYUAN · ARXIV 2609.04034 Diffusion image models produce flattened bitmaps with error-prone text and no way to edit a layer afterwards, while code-based generation gives clean layers but weak aesthetic judgment. The paper proposes a coding-agent paradigm in which a vision-language model acts as the creative brain and emits code, keeping layout control and editable layers. The 12-author work comes from Tencent Hunyuan.
  4. 4 The Missing Temporal Link: Temporal Context Routing for Script-Driven Audio-Video Generation 33 UPVOTES · YICHEN LIU ET AL. · ARXIV 2609.02367 Joint audio-video generators have gotten good at lip and scene synchronization but still offer little control over when a shot cuts or a line of dialogue lands, which breaks script-driven production. The authors show that timing written into a structured prompt survives only in the text representation and never reaches the generator's temporal axis. Their temporal context routing carries it through explicitly.
  5. 5 Last Translation Benchmark 25 UPVOTES · VILÉM ZOUHAR ET AL. · ARXIV 2609.04173 A 244-author collaboration argues that standard machine translation benchmarks are approaching saturation while automatic metrics remain unreliable, open to reward hacking and unactionable, and even gold human evaluation lacks reproducibility and scale. The paper offers a benchmark and evaluation method built to test the limits of current models and surface specific failure cases rather than emit one number.