xClean Tools

AI Papers — Daily Top 5

UPDATED 2026-11-03 01:12 PST

2026-09-04 · FRIDAY · 13:05 PDT

  1. 1 Compile by Training: Turning Natural-Language Specifications into Local Neural Functions 230 UPVOTES · YUNTIAN DENG ET AL. · UNIVERSITY OF WATERLOO · ARXIV 2609.04199 Researchers at the University of Waterloo propose compile by training, which turns a natural-language specification into a reusable neural function. Teacher models generate task-specific examples at compile time to train a small adapter for a compact interpreter, so the result runs without the teachers and can be stored, versioned and composed like ordinary software. It leads the day's papers with 230 votes.
  2. 2 Terminal-Universe: Turning Agent Trajectories into Scalable Terminal Environments 209 UPVOTES · JIE WU ET AL. · QWEN · ARXIV 2609.04148 A Qwen team reconstructs executable terminal environments from the tool-execution history recorded in existing agent trajectories. The authors argue that post-training needs environments rather than trajectories, since each environment can be re-queried into many verifiable tasks and returns execution feedback, while a trajectory is a single frozen demonstration. The paper drew 209 votes.
  3. 3 Knowing When Not to Reuse: Conditional Experience Transfer in Autonomous LLM Post-Training 139 UPVOTES · TINGYUN LI ET AL. · ARXIV 2608.26730 This paper asks which past update evidence still applies once later training has changed the parent model. In autonomous post-training systems that propose updates, train candidates and select from evaluation feedback, an update's effect depends on its parent, data and training stage, so treating earlier success as context-free permission wastes compute. It collected 139 votes.
  4. 4 Random Attention: Rethinking KV Cache Eviction for Efficient Reasoning 108 UPVOTES · HENG WANG ET AL. · SALESFORCE AI RESEARCH · ARXIV 2609.03430 Salesforce AI Research reports that the scoring signal behind KV cache eviction contributes almost nothing. Their Random Attention keeps the prompt and evicts uniformly at random within each attention head, computing no score at all, and across four models and six reasoning tasks it matches the strongest prior evictor. The paper drew 108 votes.
  5. 5 It Takes Two to Match: Co-Evolving Generative Retriever with Reinforcement Learning 69 UPVOTES · RUNPENG DAI ET AL. · APPLE · ARXIV 2609.00638 Apple researchers present CoGR, which trains language models to build retrieval representations on both the query side and the item side rather than using generation only to expand queries. Retrieval is the first stage of search and advertising systems, selecting candidates for downstream ranking and auction. The paper drew 69 votes.

2026-09-03 · THURSDAY · 13:06 PDT

  1. 1 HarnessDev: Can LLMs Create and Evolve Their Own Agent Harness? 178 UPVOTES · YUHAO WU ET AL. · BYTEDANCE SEED · ARXIV 2609.01437 ByteDance Seed introduces HarnessDev, a benchmark that evaluates models on building and evolving the agent harness itself rather than on task outputs. The unit of evaluation becomes runnable infrastructure, split across two stages. The premise is that changing the harness while holding weights fixed can substantially alter task performance, an ability current evaluations leave underexplored.
  2. 2 Language Models Can Control Their Own Attention 50 UPVOTES · NAMGYU HO ET AL. · KAIST AI · ARXIV 2609.02737 KAIST AI proposes an intrinsic route to sparse attention: rather than pre-selecting context tokens with lightweight external proxy scores, the model itself signals which parts of the context matter. The target is the cost of global attention layers reading the full KV cache at every step, which proxy scoring still pays at O(N) per step.
  3. 3 H3-World: Turning Language Understanding into World Control 47 UPVOTES · DANZE CHEN ET AL. · ARXIV 2609.01560 H3-World turns the 33B MiniMax-H3 video generator into an interactive world model without adding dedicated action modules. Each action is represented as a structured language instruction, sharpening the generator's existing zero-shot control of character behavior and camera motion into precise, temporally grounded control.
  4. 4 ZimaBlue: Evolving Generalizable World Action Models through Scalable Video Pre-training 43 UPVOTES · XIONGHAO WU ET AL. · JOY FUTURE ACADEMY · ARXIV 2609.00188 ZimaBlue learns generalizable world action models for robotic manipulation from egocentric video instead of action-labeled robot trajectories. It addresses a scaling gap: robust generalization needs broad physical experience, yet labeled trajectories are expensive and narrow. Action-free video supplies contact dynamics, tool use and long-horizon behavior across diverse settings.
  5. 5 From Production Traffic to Post-Training: Building a Self-Hosted LLM That Covers the Corporate Request Mix 33 UPVOTES · OLGA TSYMBOI ET AL. · T-TECH · ARXIV 2609.01572 T-Tech reports consolidating traffic from more than 200 internal applications onto a single self-hosted model. Quality gaps were found through production error analysis along three axes - instruction following, function calling and internal task distribution - and tracked with offline benchmarks stratified to production traffic. The work targets the GPU fragmentation that data-residency rules create.

2026-09-02 · WEDNESDAY · 01:04 PDT

  1. 1 StudentSim: Training LLM-based Student Simulators 220 UPVOTES · KE YANG ET AL. · MICROSOFT RESEARCH · ARXIV 2609.01591 Microsoft Research introduced StudentSim, a training framework for LLM-based simulators that stand in for real learners when evaluating AI tutors. It targets a gap between state-tracking models, which fit student behavior but handle explanations and corrections poorly, and LLM role-play, which follows guidance fluently without matching the imitated student's competence. It is the day's most upvoted paper.
  2. 2 DreamX-Creator: Democratizing Native Audio-Video Generation at 2K Resolution 93 UPVOTES · JIASHU ZHU ET AL. · AMAP-ML · ARXIV 2608.31106 AMAP-ML presented DreamX-Creator 1.0, a compact system built on a 7B generator that denoises audio and video jointly instead of adding sound in a separate stage. Conditioned on a first frame and a text prompt, the two streams run independently through the first half of the network and are coupled in the latter half by gated cross-modal attention.
  3. 3 GenFirst: Generation Before Reconstruction for Stable End-to-End Latent Generative Modeling 60 UPVOTES · GUANGTING ZHENG ET AL. · BYTEDANCE SEED · ARXIV 2608.29335 ByteDance Seed revisited end-to-end latent generative modeling, where the autoencoder and the generator are trained together rather than fitting a generator on a frozen, reconstruction-optimized latent space. Analyzing how different objectives shape that space, the authors target the latent collapse and the generation-reconstruction conflict that have made joint training unstable.
  4. 4 Qwen-Drive-1.0: An Initial Step towards a Vision-Language Foundation Model for Autonomous Driving 56 UPVOTES · XIN ZHOU ET AL. · QWEN · ARXIV 2609.00111 Qwen released Qwen-Drive-1.0, a vision-language foundation model for autonomous driving that keeps the pretrained VLM architecture and folds 3D perception, visual question answering and motion planning into one framework. An external bird's-eye-view head performs 3D detection, occupancy prediction and map segmentation, probing what 3D structure the shared representations expose.
  5. 5 SMELT: Scaling Laws for Compute-Matched MoE Looped Transformers 53 UPVOTES · SHAOWEN WANG ET AL. · BYTEDANCE SEED · ARXIV 2609.01343 ByteDance Seed measured looped Transformers under matched compute, holding per-token FLOPs, non-embedding parameters and KV cache constant so that extra depth is not confused with extra FLOPs. The resulting SMELT recipe loops the middle half of a sparse mixture-of-experts model's layers twice, and is scaled across four sizes up to 54B non-embedding parameters.

2026-09-01 · TUESDAY · 13:05 PDT

  1. 1 Does On-Policy Distillation Really Distill? From Noisy Teacher to Self-Improvement 90 UPVOTES · YI DING ET AL. · PURDUE UNIVERSITY · ARXIV 2608.31046 Purdue researchers measured the supervision a teacher supplies during on-policy distillation, where the teacher scores trajectories that are off-policy for it, and found substantial noise that becomes more prevalent as the teacher grows larger. The student policy converges regardless, which the authors read as pointing to self-improvement rather than transfer from the teacher.
  2. 2 Lucida: Parse, Generate, and Place for Composable Real-to-Sim Scene Modeling 65 UPVOTES · MINGHAN QIN ET AL. · BYTEDANCE SEED · ARXIV 2608.30821 ByteDance Seed presents Lucida, a pipeline that turns a capture of a real indoor room into complete, individually editable object assets arranged as they were observed. It targets the weak point of existing parse-generate-place systems, each stage of which assumes accurate instance geometry, unoccluded views and assets that match the scene. The output is aimed at robot simulation and embodied AI.
  3. 3 J-Zero: Unified Challenger--Solver--Judge Co-Evolution from Zero Data 37 UPVOTES · GYOUK CHU ET AL. · KAIST AI · ARXIV 2608.26582 KAIST AI trains a challenger, a solver and a judge together from no seed data, so a model can keep improving in domains where an answer cannot be checked automatically. The challenger raises task difficulty as the solver adapts to it, while the judge co-adapts alongside them to supply the reward signal that unverifiable tasks lack.
  4. 4 Normalized Low-Rank Adaptation 37 UPVOTES · JIALE KANG ET AL. · ARXIV 2608.31036 NoRA normalizes the down-projection matrices during LoRA training, on the observation that a zero-initialized up-projection leaves early optimization almost entirely to the down-projection. The authors report that applying the same normalization only at initialization already improves standard LoRA, making it a drop-in change to a widely used fine-tuning method.
  5. 5 StarHarness: Evolving Harnesses with Stratified Search for Enterprise Environments 35 UPVOTES · ESAKKIVEL ESAKKIRAJA ET AL. · SERVICENOW-AI · ARXIV 2608.24804 ServiceNow's StarHarness evolves the scaffolding around an agent, including task framing, tool interfaces, skills, MCP-backed providers, subagent structure and agent-loop configuration, while model weights stay fixed. It stratifies tasks by baseline failure behavior and separates proposer-visible tasks from those held back for selection and evaluation, tested on IT operations and finance benchmarks.

2026-08-31 · MONDAY · 13:06 PDT

  1. 1 LoopArena: Benchmarking Models as Runtime Controllers for Loop Engineering 82 UPVOTES · YI WANG ET AL. · AMAP-ML · ARXIV 2608.28281 A seven-author team at AMAP-ML introduces LoopArena, a benchmark that scores a model as the runtime controller of a coding-agent loop rather than as the coder. The authors note that the outcome of one end-to-end run cannot separate the loop's guidance from the agent's own ability, so the benchmark isolates decisions such as when to verify, where to spend budget and when to stop.
  2. 2 DART-SD: Diamond-topology Aware Retrieval and Tuning for Self-Distillation of Multi-Turn Tool-Calling Agents 60 UPVOTES · HANGRUI XU ET AL. · BYTEDANCE · ARXIV 2608.18524 ByteDance researchers propose DART-SD, a self-distillation recipe for multi-turn tool-calling agents. They argue that imitating full-length trajectories collapses the combinatorial lattice formed by order-independent sub-goals, indiscriminately penalizing valid alternative orderings and degrading policy diversity. DART-SD instead pairs topology-aware retrieval with tuning to keep that structure.
  3. 3 Puro-2B: Poor Lab's Qwen2-1.5B Trained on RTX 5090 within $5090 26 UPVOTES · KAIRONG LUO ET AL. · PACMAN GROUP, TSINGHUA UNIVERSITY · ARXIV 2608.27370 An 11-author team from Tsinghua University's PACMAN Group publishes an open pretraining recipe and the Puro-2B models it produces, trained on an RTX 5090 for under 5,090 dollars. The report notes that training Llama-3.2-3B costs over 1.5 million dollars and reproducing SmolLM3-3B over 700,000, and argues a cost-efficient, hardware-accessible recipe has been the missing piece for academic and open-source groups.
  4. 4 Act with Intent: Distilling Behavior Intent for Vision-Language-Action Models 24 UPVOTES · SANGOH LEE ET AL. · POHANG UNIVERSITY OF SCIENCE AND TECHNOLOGY · ARXIV 2608.23478 Three researchers at POSTECH propose Intention Distillation for vision-language-action models, whose action decoders are still trained largely by behavior cloning. That supervision records which motor command was demonstrated but leaves the local objective implicit, and the authors argue future-based signals such as frames or trajectories capture particular realizations rather than the shared aim of the behavior.
  5. 5 LayerRecall: A State-Conditioned Memory Router for Long-Horizon Consistency in Video Generation 24 UPVOTES · YIXUAN DING ET AL. · ZHEJIANG UNIVERSITY · ARXIV 2608.28460 A Zhejiang University team introduces LayerRecall, a memory router for autoregressive video diffusion, which generates long clips chunk by chunk from a bounded recent context and evicts the historical cues needed when a subject or scene reappears. Their analysis finds video DiT layers differ in their preference for current, recent and distant context, so the router decides both what to retrieve and where to use it.

2026-08-29 · SATURDAY · 13:04 PDT

  1. 1 Zero-WAM: In-Context World-Action Modeling from Human Videos for Open-Ended Task Generalization 17 UPVOTES · JIAMING ZHOU ET AL. · ROBBYANT RESEARCH · ARXIV 2608.26103 Zero-WAM carries in-context learning over from language models to robot manipulation, using a human video rather than a sentence as the specification of a new task. The authors argue video is the natural specification for manipulation because it supplies visual cues that language leaves out. The policy then executes tasks never seen in training without any parameter update.
  2. 2 WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution 11 UPVOTES · LIYAN TANG ET AL. · GOOGLE · ARXIV 2608.27454 Google proposes WikiSkill, a framework that evolves an agent's skill library alongside a persistent knowledge base rather than in isolation. It targets a specific gap in automatic skill discovery: the insights that guide how a skill develops stay scattered through optimization histories and are rarely reused across iterations. WikiSkill keeps raw execution experience and accumulated knowledge in separate layers.
  3. 3 Procedura: Agentic 3D Modeling with Procedural Control 10 UPVOTES · YOUTIAN LIN ET AL. · NANJING UNIVERSITY · ARXIV 2608.26238 Procedura treats a 3D shape as code, scaling an LLM's coding ability to write an object as a procedural assembly whose named parts are joined by typed, machine-checkable mates. The team from Nanjing University aims at three weaknesses of dense meshes from native 3D generators: soft edges where an object should be sharp, no part decomposition, and no parameter a user can edit.
  4. 4 CritICL: Inference-Time Weak-to-Strong Generalization from Small Language Model Failure Modes 8 UPVOTES · YUFAN WU ET AL. · ARXIV 2608.27455 CritICL improves reasoning at inference time without the repeated generation or external verification that inference-time scaling usually requires. Its premise is that failure modes are structured and recur across model scales within one family, so the mistakes of a small model become guidance for a larger one. The authors present failures as a source of signal rather than output to discard.
  5. 5 Magpie: Real-Time World Renderer for Interactive Games 7 UPVOTES · XIAOYU ZHAN ET AL. · ARXIV 2608.27168 Magpie is a real-time generative world renderer built for games rather than linear media. The authors note that video foundation models now reshape film and video work, but a game additionally needs stable, reproducible rules, object states and interaction outcomes on top of continuous imagery. The target is the cost of asset production, which stretches prototype development cycles.

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.