xClean Tools

AI Papers — Daily Top 5

UPDATED 2026-11-03 01:12 PST

2026-09-23 · WEDNESDAY · 13:11 PDT

  1. 1 The Tasteful Agent: Measuring and Improving Taste in Long-Horizon Tasks 111 UPVOTES · WENBO PAN ET AL. · MICROSOFT · ARXIV 2609.25804 Microsoft researchers define an agent's taste as its ability to choose well at forks in long-horizon work, and release Taste-Bench to measure it. Questions are mined automatically from parallel attempts and detours in real trajectories, and the model must pick a direction without seeing what follows. The best frontier model answers 59.7 percent correctly, and a larger reasoning budget does not help.
  2. 2 GAE: Learning a Geometry-Native Latent Space for 3D-Consistent World Generation 38 UPVOTES · JIAHAO LU ET AL. · ARC LAB, TENCENT · ARXIV 2609.24981 Tencent's ARC Lab argues that 3D inconsistency in video generation is a representation problem, and reparameterizes a geometry foundation model's features into a compact latent decodable to appearance, depth, cameras and point maps. Holding the generator and training protocol fixed, swapping in this latent cuts FVD by 12.7 percent on RealEstate10K and 23.1 percent on DL3DV, and halves camera-trajectory error.
  3. 3 All-in-One Multilingual Scene Text Recognition with Script-aware Mixture-of-Experts 35 UPVOTES · XINGSONG YE ET AL. · FUDAN UNIVERSITY · ARXIV 2609.24058 Fudan University researchers built ScriptMoE, a scene text recognizer that shares one visual encoder and replaces the dense decoder with a sparse mixture of experts, routing each image to its top-2 script-aligned experts. It trains on TextMuSS-10M, a synthetic set spanning 10 scripts and 229 languages. Swapping it into PP-OCRv5 lifts multilingual F1 from 65.71 to 80.89, past the best VLM at far fewer parameters.
  4. 4 onPanda: Efficient Annotation of On-Policy Alignment Data for LLMs and Agents via Token-Level Correction 35 UPVOTES · LEI YANG ET AL. · STEPFUN · ARXIV 2609.24983 StepFun presented onPanda, an interactive tool for annotating LLM alignment data and agent trajectories, built around token-level correction. The annotator finds the first inappropriate token, then picks a substitute from the model's own candidates or types a replacement; the system truncates everything after it and regenerates from the corrected prefix.
  5. 5 One to More, More to One: Category-Aware Iterative Expert Training for Software Engineering Agents 31 UPVOTES · JIE ZHAO ET AL. · ALIBABA · ARXIV 2609.23377 Alibaba researchers target the category see-saw in repository-level software engineering, where pooled agentic reinforcement learning lifts some task categories while regressing others. They train per-category experts, then merge them into one deployable student by label-routed multi-teacher distillation. It resolves 58.04 percent on Pro-618 and 59.00 percent on SWE-bench Multilingual.

2026-09-22 · TUESDAY · 13:10 PDT

  1. 1 GameHorizon Suite: Multi-Horizon Data and Evaluation in Gameplay 110 UPVOTES · YIRAN WANG ET AL. · TENCENT · ARXIV 2609.25001 Tencent introduces GameHorizon, a data and evaluation suite that measures gameplay ability at different time horizons. It pairs an automated pipeline for multi-horizon instruction annotation with 5,000 hours of AAA recordings from 21 games collected by 100 expert players, plus a benchmark with reproducible offline questions and stepwise online tests. The team evaluated 47 models over more than a million invocations.
  2. 2 Transferring the Intelligence of VLMs to Robotic Control 101 UPVOTES · MENG-HAO GUO ET AL. · ARXIV 2609.22966 RoboDawn hands robot control to an agentic vision-language model through a compact set of discrete translation, rotation and gripper commands, closing the loop on visual observations. With no task-specific robot training it reaches 53.2 percent success on RoboTwin 2.0 C2R, rising to 73.6 percent with one in-context demonstration against 46.0 percent for the pi 0.5 baseline. It also runs on a real Franka arm.
  3. 3 OmniEdu: Open Foundation Models for Learning and Teaching 66 UPVOTES · HAO LIANG ET AL. · PEKING UNIVERSITY · ARXIV 2609.23088 OmniEdu is an open family of K-12 education models at 4B, 9B and 27B. Its instruction-tuning corpus draws on more than 100 educational and general sources organized around subject competence, curriculum grounding, diagnostic reasoning and pedagogical scaffolding, yielding 69,999 examples and 15.96M supervised response tokens. OmniEdu-27B reaches 63.12 percent exact match on K12-Bench and 86.95 percent on EDUMATH.
  4. 4 Document Retrieval-Aware Chunking (D-RAC): Universal Retrieval-Aware Ingestion of Enterprise Documents via PDF Normalization and Multimodal Markdown Conversion 48 UPVOTES · UDAY ALLU ET AL. · YELLOW.AI · ARXIV 2609.24220 Yellow.ai extends its retrieval-aware chunking to any document format by first normalizing input to PDF, then using one multimodal LLM pass to turn rendered pages into retrieval-optimized Markdown that keeps heading hierarchy and rewrites tables as self-contained prose. Chunk planning then runs over identifiers rather than text, cutting chunking-stage output tokens by 95.7 percent versus agentic chunking.
  5. 5 VideoGen-Agent: Reinforcing Video Generation Agents 33 UPVOTES · BINXU LI ET AL. · PRINCETON UNIVERSITY · ARXIV 2609.24997 VideoGen-Agent is a multimodal agent trained with multitask agentic reinforcement learning to coordinate augmentation, generation and verification tools across turns. On VABench, a held-out set of 600 prompts covering procedural knowledge, identity, physical consistency and multi-shot structure, it lifts its base text-to-video generator from 56.5 to 75.6, and to 86.1 once the generation tools are upgraded.

2026-09-21 · MONDAY · 13:10 PDT

  1. 1 CodeMidas: Scaling Agentic Coding RL Environments from Code Itself 83 UPVOTES · BOWEN YE ET AL. · XIAOMI MIMO · ARXIV 2609.22068 CodeMidas turns functionality already implemented in open-source codebases into executable RL environments for coding agents, using source code as its only task-specific input. Agents write behavioral specs and execution-grounded tests, then filter tasks through repeated rollouts, yielding 5,545 tasks from 3,185 repos in 23 languages; GRPO training lifted MiMo-V2.5 on all five benchmarks, by 17% on ProgramBench.
  2. 2 IntBMoE: Integrating Block-Level Conditioning into Expert Composition for Full-Participation Mixture-of-Experts 33 UPVOTES · RAN CHENG ET AL. · ARXIV 2609.21346 IntBMoE is a mixture-of-experts design in which every expert contributes to each token while compute and memory stay bounded. A hypernetwork merges each layer's expert pool into composed experts drawn from a small learned codebook of blocks, and a router sends each token to only a few blocks. It beat sparse and dense MoE baselines on image classification and lifted click-through 2.4% in AMap's recommender.
  3. 3 OmniVChat: Synthesizing, Benchmarking, and Training for Native Audio-Visual Dialogue 29 UPVOTES · HAOLIN HE ET AL. · QWEN · ARXIV 2609.21465 OmniVChat, from Qwen, defines native audio-visual dialogue: an omni model takes a user's audio and video directly, with no text question, captions or speech recognition, and replies in text. The team built OmniVChat-Studio, a multi-agent dialogue synthesizer, plus a five-category benchmark and an RL reward for correctness, efficiency and style that improved Qwen3-Omni-Instruct, even on human-recorded tests.
  4. 4 Designer-RSI: Evolving Procedural Memory from User Traffic for Agentic Graphic Design 19 UPVOTES · HONGYANG DU ET AL. · ADOBE · ARXIV 2609.22086 Designer-RSI lets a frozen frontier model that operates professional design software through more than 230 tools improve from real use by growing an external memory of natural-language design procedures instead of updating weights. Over five rounds on 1,406 real user briefs, the skill bank grew from 76 to 139 skills and Claude Sonnet 4's execution success on GenEval2 rose from 72.7% to 99.3%, with no human labels.
  5. 5 BI-Agent and BI-Bench: Towards Automating End-to-End Business Intelligence 12 UPVOTES · CHUXUAN HU ET AL. · MICROSOFT RESEARCH · ARXIV 2609.20886 Microsoft Research built BI-Bench from question-answer pairs in real business intelligence dashboards to test whether LLMs can answer BI questions end to end, without users first preparing tables, transforms and joins. Frontier LLMs scored below 50%. Its tool-augmented BI-Agent, which splits the work into search, join and transform steps, gained up to 40 points; a post-trained version gained up to 30.

2026-09-19 · SATURDAY · 13:09 PDT

  1. 1 Verifiable Social Reasoning for LLM Assistants 33 UPVOTES · AMIR TAUBENFELD ET AL. · GOOGLE · ARXIV 2609.17496 Google's Fuse is a multi-agent simulation that tests how LLM assistants reason about social situations relayed by a user: an agent with a hidden motive interacts with others, then the user asks the assistant to infer the motive, yielding verifiable ground truth. Across 12 LLMs, user relaying and biased framing hurt accuracy, and longer chats did not always help. Fuse and its 21k-example dataset are open-source.
  2. 2 RetireOPD: Self-Retiring On-Policy Distillation for Agentic Reinforcement Learning 26 UPVOTES · YAN YU ET AL. · ARXIV 2609.20784 RetireOPD trains multi-turn agents with reinforcement learning plus on-policy distillation from a skill-conditioned teacher, which the student retires on its own once their gap stops shrinking and it reaches a target share of the teacher's success. On 1.5B to 7B Qwen2.5 models, it lifted ALFWorld success over an RL baseline by 14.1% to 18.8% and WebShop accuracy by 11.8% to 19.0%, beating its teacher every time.
  3. 3 Reflect, Revise, Reuse: Training-Free Skill Evolution for GUI Agents 25 UPVOTES · BOFAN CHEN ET AL. · ARXIV 2609.17653 EvoSkill-GUI lets GUI agents improve skills during deployment without training: each skill is a multi-file package of plans, backup localization, recovery rules and past failures, updated in a reflect-revise-reuse loop where an isolated critic diagnoses failed runs and the executor edits specific skill files. It lifted several base models by up to 16.2% on MobileWorld, 6.0% on AndroidWorld and 10.5% on OSWorld.
  4. 4 Self-Evolving Search Index 25 UPVOTES · SANGAM LEE ET AL. · ARXIV 2609.19656 SELF-INDEX lets a search index adapt to its retrieval environment without human help: an Optimizer diagnoses retrieval shortfalls, revises the responsible index keys and validates each change, while a Query Simulator probes demands beyond existing queries. It beat prior index optimization methods across corpora and retrievers, and the gains carried over to search agents and agent memory systems.
  5. 5 Video DeltaNet: A Video-Native Hybrid Attention for Livestream Video Generation 23 UPVOTES · HAOCHENG XI ET AL. · ARXIV 2609.20744 Video DeltaNet speeds up video diffusion by pairing local Softmax attention with a bidirectional linear memory updated once per frame, and a staged teacher-alignment recipe adds the new path to pretrained models. Built on MiniMax H3 with eight-step distillation and an optimized SGLang stack, it denoises a 14.3-second 768p video in 6.70 seconds on eight Nvidia B200 GPUs, 14.5x faster than the 50-step dense baseline.

2026-09-18 · FRIDAY · 13:12 PDT

  1. 1 DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression 53 UPVOTES · DEEPSEEK-AI ET AL. · DEEPSEEK · ARXIV 2609.19969 DeepSeek's technical report introduces DeepSeek-V4.1-Flash, a multimodal mixture-of-experts model with 552B backbone parameters and contexts of up to one million tokens. Cross-layer KV cache reuse and FP4 caching cut its in-HBM KV cache to 890 bytes per token, about a quarter of DeepSeek-V4-Flash's, and the model activates 16B parameters per token when decoding and 8B during prefill.
  2. 2 SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness 42 UPVOTES · HAOZHE LIU ET AL. · NVIDIA · ARXIV 2609.20519 NVIDIA researchers ran auto-research loops across many environments to improve coding-agent harnesses, keeping four mechanisms for action execution, context compaction, observation handling and delegated reading. On the 51-task EdgeBench, the resulting SoL-Pi harness matched the Pi harness with GPT-5.6 Sol and Opus 5 while cutting token traffic by 44.7 to 49.0 percent and API cost by about a third.
  3. 3 VC-Attention: Value Smoothing and Softmax Casting for Low-bit Attention 35 UPVOTES · XINGYANG LI ET AL. · NUNCHUX AI · ARXIV 2609.15810 VC-Attention is a training-free low-bit attention method for video diffusion transformers that tackles both accuracy loss from value outliers and the softmax bottleneck. It clusters value tokens so each hardware block quantizes well and maps scores straight to FP8 probability codes with one fused multiply-add, speeding the attention kernel 1.46x to 1.59x over BF16 FlashAttention-4 on Blackwell and Hopper GPUs.
  4. 4 EvolveTrade: Experience-Driven Policy Refinement for Self-Evolving LLM Trading Agents 31 UPVOTES · SEHEE KIM ET AL. · KAIST AI · ARXIV 2609.17632 Researchers at KAIST AI propose EvolveTrade, which treats a trading agent's system prompt as a policy that a separate Policy Agent rewrites after each interval using past decision traces and realized portfolio returns, with the backbone LLM kept fixed. Across several market regimes and two LLM backbones, the evolved policies improved Sharpe ratio and cumulative return over fixed-policy baselines in most settings.
  5. 5 Zing-0.5: Toward Playable Worlds with Real-Time Joint Action and Text Control 29 UPVOTES · MINGYANG CHEN ET AL. · ARXIV 2609.17909 Zing-0.5 is a 5B-parameter autoregressive world model built for playability, letting users explore a generated world with the keyboard while steering events through live text prompts. It streams 832x480 video at 24 frames per second using four-step generation, at an estimated server cost of about 0.009 dollars per stream-minute, and the team is releasing the weights, inference code and an SGLang serving stack.

2026-09-17 · THURSDAY · 13:07 PDT

  1. 1 Continual Learning Mechanisms Compose for Long-Horizon Memorization 294 UPVOTES · ZHEYUAN ZHANG ET AL. · JOHNS HOPKINS UNIVERSITY · ARXIV 2609.06986 The authors introduce long-horizon memorization, a setting where a model learns 100 query-answer tasks through continual fine-tuning without keeping earlier examples or receiving task identifiers at inference. No single continual learning mechanism they evaluate holds retention across that horizon. Combining mechanisms that address different sources of forgetting does better than any one alone.
  2. 2 AI for Games in the Foundation Model Era 113 UPVOTES · MENG LUO ET AL. · NATIONAL UNIVERSITY OF SINGAPORE · ARXIV 2609.16679 This survey maps how foundation models and learned game-world models are being applied across the game lifecycle. It sorts the literature into six roles defined by what the AI output is used for, from playing and acting to modeling players and evaluating artifacts. The authors argue these strands developed separately, obscuring which capabilities transfer between games and engines.
  3. 3 StepAudio 3 Realtime Technical Report 99 UPVOTES · BIN LIN ET AL. · STEPFUN · ARXIV 2609.14005 StepFun describes StepAudio 3 Realtime, an audio-language foundation model organized around a continuous listen-converse-think-act loop. A duplex component models both audio streams together to handle pauses, backchannels and interruptions. A Think-While-Speaking scheme runs private reasoning in parallel with spoken delivery to keep deliberation from adding latency.
  4. 4 LimiX-2: A Contextual Mechanism Network Towards General Structured-Data Intelligence 87 UPVOTES · XINGXUAN ZHANG ET AL. · STABLE AI · ARXIV 2609.17488 LimiX-2 scales the LimiX family along the authors' previously published scaling laws, using Contextual Mechanism Networks pretrained with context-conditional masked modeling. Instead of centering on predicting a target from context as conventional tabular prior-fitted networks do, it is built around learning the joint distribution of features and labels given the context.
  5. 5 ScienceIDE: Turning World's Scientific Codebase into Agent Learnable Environments 69 UPVOTES · HEJIA GENG ET AL. · PHAI LABS · ARXIV 2609.19134 ScienceIDE is infrastructure for converting scientific code repositories into programmable environments that agents can learn in. Guided by expert-defined cases and acceptance criteria, agents turn a repository into an executable environment. The authors call the target problem a scientific experience bottleneck, created by fragmented toolchains and implicit domain conventions.

2026-09-14 · MONDAY · 01:10 PDT

  1. 1 Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation 43 UPVOTES · KOUTIAN WU ET AL. · CARNEGIE MELLON UNIVERSITY · ARXIV 2609.11115 Benchmark Radar is a living database and search engine for AI benchmarks, covering LLM evaluation, agentic and tool-use tests, coding, reasoning, safety and domain-specific evaluations. It pairs daily discovery of benchmark papers, repositories, datasets and releases with a searchable catalog. It aims to help users locate datasets and code and understand the settings behind reported scores.
  2. 2 SAS: Simple Attention Sparsification via End-to-End Optimization of Context Ranking 41 UPVOTES · ZHIWEI LI ET AL. · TENCENT HUNYUAN · ARXIV 2609.13141 Tencent Hunyuan researchers present SAS, a post-training method that cuts the quadratic attention cost of pretrained Transformers by picking a small set of tokens or blocks for each query. Existing trainable selectors apply hard Top-K selection that blocks gradients from the language modeling loss, so they usually distill dense attention patterns instead. SAS optimizes the context ranking end to end.
  3. 3 Breaking the Vision-Action Shortcut: Latent Interface Training for Generalizable Robotics Foundation Models 37 UPVOTES · JIANMAN LIN ET AL. · NATIONAL UNIVERSITY OF SINGAPORE · ARXIV 2609.12641 Latent Interface Training aims to keep robot foundation models reliable under visual distribution shifts. Policies built on pretrained visual features can exploit task-irrelevant cues that correlate with demonstrated actions, a shortcut that breaks when those correlations change. The method constrains how visual information is used for action generation while preserving task-relevant spatial detail.
  4. 4 COBRA-Skills: Contextual Bandit-Guided Evolution for Agent Skill Optimization 24 UPVOTES · PINGCHEN LU ET AL. · CHINESE UNIVERSITY OF HONG KONG, SHENZHEN · ARXIV 2609.11682 COBRA-Skills optimizes the reusable skills LLM agents distill from prior tasks, a job existing methods often handle with costly execution-based evaluation and substantial task data. It casts skill optimization as budgeted sequential optimization over an evolving candidate space. A contextual bandit directs evaluations to promising or informative candidates, while evidence-grounded evolution refines the skills.
  5. 5 PLC-DPO: Posterior Label Correction in Noisy and Ambiguous Preference Optimization 22 UPVOTES · BORYEONG CHO ET AL. · KAIST AI · ARXIV 2608.30597 KAIST AI researchers propose PLC-DPO, a version of Direct Preference Optimization built for preference data with reversed, weak or ambiguous labels. Standard DPO assumes every observed preference is reliable, so noisy pairs can cause harmful policy updates. PLC-DPO routes each pair's training signal as a clean, flipped or tied case, using the calibrated policy-reference margin as online evidence for the correction.