https://tools.xclean.dev/papers Top AI papers (2026-09-05) from the Hugging Face community: RoboTok: An Internet-Scale Data Engine for Human Demonstration Retrieval and Dexterous Manipulation Learning and more, each with a plain-language summary. en Sat, 05 Sep 2026 20:05:00 GMT https://huggingface.co/papers/2609.03199https://huggingface.co/papers/2609.03199#2026-09-05Sat, 05 Sep 2026 00:00:00 GMTRoboTok 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2609.04201https://huggingface.co/papers/2609.04201#2026-09-05Sat, 05 Sep 2026 00:00:00 GMTOnline 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2609.04034https://huggingface.co/papers/2609.04034#2026-09-05Sat, 05 Sep 2026 00:00:00 GMTDiffusion 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2609.02367https://huggingface.co/papers/2609.02367#2026-09-05Sat, 05 Sep 2026 00:00:00 GMTJoint 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2609.04173https://huggingface.co/papers/2609.04173#2026-09-05Sat, 05 Sep 2026 00:00:00 GMTA 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2609.04199https://huggingface.co/papers/2609.04199#2026-09-04Fri, 04 Sep 2026 00:00:00 GMTResearchers 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2609.04148https://huggingface.co/papers/2609.04148#2026-09-04Fri, 04 Sep 2026 00:00:00 GMTA 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.26730https://huggingface.co/papers/2608.26730#2026-09-04Fri, 04 Sep 2026 00:00:00 GMTThis 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2609.03430https://huggingface.co/papers/2609.03430#2026-09-04Fri, 04 Sep 2026 00:00:00 GMTSalesforce 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2609.00638https://huggingface.co/papers/2609.00638#2026-09-04Fri, 04 Sep 2026 00:00:00 GMTApple 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2609.01437https://huggingface.co/papers/2609.01437#2026-09-03Thu, 03 Sep 2026 00:00:00 GMTByteDance 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2609.02737https://huggingface.co/papers/2609.02737#2026-09-03Thu, 03 Sep 2026 00:00:00 GMTKAIST 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2609.01560https://huggingface.co/papers/2609.01560#2026-09-03Thu, 03 Sep 2026 00:00:00 GMTH3-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.AI Papers — Daily Top 5 https://huggingface.co/papers/2609.00188https://huggingface.co/papers/2609.00188#2026-09-03Thu, 03 Sep 2026 00:00:00 GMTZimaBlue 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2609.01572https://huggingface.co/papers/2609.01572#2026-09-03Thu, 03 Sep 2026 00:00:00 GMTT-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.AI Papers — Daily Top 5 https://huggingface.co/papers/2609.01591https://huggingface.co/papers/2609.01591#2026-09-02Wed, 02 Sep 2026 00:00:00 GMTMicrosoft 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.31106https://huggingface.co/papers/2608.31106#2026-09-02Wed, 02 Sep 2026 00:00:00 GMTAMAP-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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.29335https://huggingface.co/papers/2608.29335#2026-09-02Wed, 02 Sep 2026 00:00:00 GMTByteDance 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2609.00111https://huggingface.co/papers/2609.00111#2026-09-02Wed, 02 Sep 2026 00:00:00 GMTQwen 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2609.01343https://huggingface.co/papers/2609.01343#2026-09-02Wed, 02 Sep 2026 00:00:00 GMTByteDance 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.31046https://huggingface.co/papers/2608.31046#2026-09-01Tue, 01 Sep 2026 00:00:00 GMTPurdue 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.30821https://huggingface.co/papers/2608.30821#2026-09-01Tue, 01 Sep 2026 00:00:00 GMTByteDance 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.26582https://huggingface.co/papers/2608.26582#2026-09-01Tue, 01 Sep 2026 00:00:00 GMTKAIST 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.31036https://huggingface.co/papers/2608.31036#2026-09-01Tue, 01 Sep 2026 00:00:00 GMTNoRA 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.24804https://huggingface.co/papers/2608.24804#2026-09-01Tue, 01 Sep 2026 00:00:00 GMTServiceNow'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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.28281https://huggingface.co/papers/2608.28281#2026-08-31Mon, 31 Aug 2026 00:00:00 GMTA 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.18524https://huggingface.co/papers/2608.18524#2026-08-31Mon, 31 Aug 2026 00:00:00 GMTByteDance 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.27370https://huggingface.co/papers/2608.27370#2026-08-31Mon, 31 Aug 2026 00:00:00 GMTAn 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.23478https://huggingface.co/papers/2608.23478#2026-08-31Mon, 31 Aug 2026 00:00:00 GMTThree 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.28460https://huggingface.co/papers/2608.28460#2026-08-31Mon, 31 Aug 2026 00:00:00 GMTA 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.26103https://huggingface.co/papers/2608.26103#2026-08-29Sat, 29 Aug 2026 00:00:00 GMTZero-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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.27454https://huggingface.co/papers/2608.27454#2026-08-29Sat, 29 Aug 2026 00:00:00 GMTGoogle 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.26238https://huggingface.co/papers/2608.26238#2026-08-29Sat, 29 Aug 2026 00:00:00 GMTProcedura 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.27455https://huggingface.co/papers/2608.27455#2026-08-29Sat, 29 Aug 2026 00:00:00 GMTCritICL 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.27168https://huggingface.co/papers/2608.27168#2026-08-29Sat, 29 Aug 2026 00:00:00 GMTMagpie 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.25518https://huggingface.co/papers/2608.25518#2026-08-28Fri, 28 Aug 2026 00:00:00 GMTThe 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.27456https://huggingface.co/papers/2608.27456#2026-08-28Fri, 28 Aug 2026 00:00:00 GMTUrbanGround 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.27448https://huggingface.co/papers/2608.27448#2026-08-28Fri, 28 Aug 2026 00:00:00 GMTPost-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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.26872https://huggingface.co/papers/2608.26872#2026-08-28Fri, 28 Aug 2026 00:00:00 GMTOn-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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.27260https://huggingface.co/papers/2608.27260#2026-08-28Fri, 28 Aug 2026 00:00:00 GMTWork 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.19583https://huggingface.co/papers/2608.19583#2026-08-27Thu, 27 Aug 2026 00:00:00 GMTA 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.25593https://huggingface.co/papers/2608.25593#2026-08-27Thu, 27 Aug 2026 00:00:00 GMTResearchers 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.23181https://huggingface.co/papers/2608.23181#2026-08-27Thu, 27 Aug 2026 00:00:00 GMTA 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.24987https://huggingface.co/papers/2608.24987#2026-08-27Thu, 27 Aug 2026 00:00:00 GMTA 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.23318https://huggingface.co/papers/2608.23318#2026-08-27Thu, 27 Aug 2026 00:00:00 GMTZhejiang 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.15875https://huggingface.co/papers/2608.15875#2026-08-26Wed, 26 Aug 2026 00:00:00 GMTGigaAI 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.23041https://huggingface.co/papers/2608.23041#2026-08-26Wed, 26 Aug 2026 00:00:00 GMTMicrosoft 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.24646https://huggingface.co/papers/2608.24646#2026-08-26Wed, 26 Aug 2026 00:00:00 GMTByteDance 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.21500https://huggingface.co/papers/2608.21500#2026-08-26Wed, 26 Aug 2026 00:00:00 GMTThe 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.22876https://huggingface.co/papers/2608.22876#2026-08-26Wed, 26 Aug 2026 00:00:00 GMTThe 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.20958https://huggingface.co/papers/2608.20958#2026-08-25Tue, 25 Aug 2026 00:00:00 GMTTLive-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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.20910https://huggingface.co/papers/2608.20910#2026-08-25Tue, 25 Aug 2026 00:00:00 GMTInstruction-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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.23035https://huggingface.co/papers/2608.23035#2026-08-25Tue, 25 Aug 2026 00:00:00 GMTMobilePA-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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.23552https://huggingface.co/papers/2608.23552#2026-08-25Tue, 25 Aug 2026 00:00:00 GMTPrime 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.19567https://huggingface.co/papers/2608.19567#2026-08-25Tue, 25 Aug 2026 00:00:00 GMTBlock3D 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.16425https://huggingface.co/papers/2608.16425#2026-08-24Mon, 24 Aug 2026 00:00:00 GMTParaTempo 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.12781https://huggingface.co/papers/2608.12781#2026-08-24Mon, 24 Aug 2026 00:00:00 GMTTencent 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.20886https://huggingface.co/papers/2608.20886#2026-08-24Mon, 24 Aug 2026 00:00:00 GMTEviRank 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.08676https://huggingface.co/papers/2608.08676#2026-08-24Mon, 24 Aug 2026 00:00:00 GMTUniSpace 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.18484https://huggingface.co/papers/2608.18484#2026-08-24Mon, 24 Aug 2026 00:00:00 GMTThe 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.17744https://huggingface.co/papers/2608.17744#2026-08-22Sat, 22 Aug 2026 00:00:00 GMTThree 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.08466https://huggingface.co/papers/2608.08466#2026-08-22Sat, 22 Aug 2026 00:00:00 GMTLLM 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.19891https://huggingface.co/papers/2608.19891#2026-08-22Sat, 22 Aug 2026 00:00:00 GMTDeepMind 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.20281https://huggingface.co/papers/2608.20281#2026-08-22Sat, 22 Aug 2026 00:00:00 GMTLanguage 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.13210https://huggingface.co/papers/2608.13210#2026-08-22Sat, 22 Aug 2026 00:00:00 GMTNARU 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.18580https://huggingface.co/papers/2608.18580#2026-08-21Fri, 21 Aug 2026 00:00:00 GMTTraining 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.19799https://huggingface.co/papers/2608.19799#2026-08-21Fri, 21 Aug 2026 00:00:00 GMTOpenMOSS 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.20336https://huggingface.co/papers/2608.20336#2026-08-21Fri, 21 Aug 2026 00:00:00 GMTTencent 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.20202https://huggingface.co/papers/2608.20202#2026-08-21Fri, 21 Aug 2026 00:00:00 GMTMemTrapBench 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.13120https://huggingface.co/papers/2608.13120#2026-08-21Fri, 21 Aug 2026 00:00:00 GMTAgent 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.17426https://huggingface.co/papers/2608.17426#2026-08-20Thu, 20 Aug 2026 00:00:00 GMTSemComp-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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.13558https://huggingface.co/papers/2608.13558#2026-08-20Thu, 20 Aug 2026 00:00:00 GMTOmniScientist 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.19197https://huggingface.co/papers/2608.19197#2026-08-20Thu, 20 Aug 2026 00:00:00 GMTSPADE 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.18940https://huggingface.co/papers/2608.18940#2026-08-20Thu, 20 Aug 2026 00:00:00 GMTInsilico 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.18063https://huggingface.co/papers/2608.18063#2026-08-20Thu, 20 Aug 2026 00:00:00 GMTEDITBRIDGE 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.14036https://huggingface.co/papers/2608.14036#2026-08-19Wed, 19 Aug 2026 00:00:00 GMTA 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.17310https://huggingface.co/papers/2608.17310#2026-08-19Wed, 19 Aug 2026 00:00:00 GMTResearchers 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.17512https://huggingface.co/papers/2608.17512#2026-08-19Wed, 19 Aug 2026 00:00:00 GMTA 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.12313https://huggingface.co/papers/2608.12313#2026-08-19Wed, 19 Aug 2026 00:00:00 GMTThe 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.16887https://huggingface.co/papers/2608.16887#2026-08-19Wed, 19 Aug 2026 00:00:00 GMTTongyi-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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.16859https://huggingface.co/papers/2608.16859#2026-08-18Tue, 18 Aug 2026 00:00:00 GMTHarnessEval-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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.15089https://huggingface.co/papers/2608.15089#2026-08-18Tue, 18 Aug 2026 00:00:00 GMTStateM 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.15669https://huggingface.co/papers/2608.15669#2026-08-18Tue, 18 Aug 2026 00:00:00 GMTThe 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.16072https://huggingface.co/papers/2608.16072#2026-08-18Tue, 18 Aug 2026 00:00:00 GMTThe 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.15045https://huggingface.co/papers/2608.15045#2026-08-18Tue, 18 Aug 2026 00:00:00 GMTOpenMOSS 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.14144https://huggingface.co/papers/2608.14144#2026-08-17Mon, 17 Aug 2026 00:00:00 GMTThe 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.11341https://huggingface.co/papers/2608.11341#2026-08-17Mon, 17 Aug 2026 00:00:00 GMTApodex 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.14530https://huggingface.co/papers/2608.14530#2026-08-17Mon, 17 Aug 2026 00:00:00 GMTMarionette 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.13517https://huggingface.co/papers/2608.13517#2026-08-17Mon, 17 Aug 2026 00:00:00 GMTMimir 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.14546https://huggingface.co/papers/2608.14546#2026-08-17Mon, 17 Aug 2026 00:00:00 GMTCPI-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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.11745https://huggingface.co/papers/2608.11745#2026-08-15Sat, 15 Aug 2026 00:00:00 GMTLiveAnimate 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.08888https://huggingface.co/papers/2608.08888#2026-08-15Sat, 15 Aug 2026 00:00:00 GMTMicrosoft 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.07193https://huggingface.co/papers/2608.07193#2026-08-15Sat, 15 Aug 2026 00:00:00 GMTVisual 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.13049https://huggingface.co/papers/2608.13049#2026-08-15Sat, 15 Aug 2026 00:00:00 GMTRobot 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.08020https://huggingface.co/papers/2608.08020#2026-08-15Sat, 15 Aug 2026 00:00:00 GMTPrinceton 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.13546https://huggingface.co/papers/2608.13546#2026-08-14Fri, 14 Aug 2026 00:00:00 GMTEvoke 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.07545https://huggingface.co/papers/2608.07545#2026-08-14Fri, 14 Aug 2026 00:00:00 GMTDarwinX 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.08975https://huggingface.co/papers/2608.08975#2026-08-14Fri, 14 Aug 2026 00:00:00 GMTThe 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.12743https://huggingface.co/papers/2608.12743#2026-08-14Fri, 14 Aug 2026 00:00:00 GMTThe 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.AI Papers — Daily Top 5 https://huggingface.co/papers/2608.12149https://huggingface.co/papers/2608.12149#2026-08-14Fri, 14 Aug 2026 00:00:00 GMTThis 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.AI Papers — Daily Top 5