arXiv papers recommended on Hugging Face, summarized in plain English.
AIarXiv:2610.31329
The authors present WeVisOne-VL, a vision-language model that unifies object-level and pixel-level perception through structured next-token prediction. They add a lightweight pixel decoder to the VLM via a special <OBJ> token to output instance-level pixel masks, and train it with three-stage curriculum SFT followed by mixed-task RL.
221 upvotes
AIarXiv:2610.30949
The authors present IQuest-Q1, a sparse Mixture-of-Experts agentic foundation model with 320B total parameters (15B activated). It is built through three stages: stable pre-training with a two-stage coding-weighted curriculum, fine-grained mid-training with progressive context extension, and multi-scaffold, multi-version post-training that combines SFT, domain-specialized RL, and MOPD, followed by model merging.
53 upvotes
AIarXiv:2610.30753
The authors present FilmFlow, an agentic framework that turns natural-language instructions or reference videos into explicit camera trajectories for dynamic 3D scenes. A Director agent composes high-level camera motions from concurrent atomic operations (translation, rotation, framing), supported by Interpreter, Scout, and Critic agents. They also release CinemaBench, with 150 Blender scenes, 750 text-directed tasks, and 60 reference videos.
26 upvotes
RoboticsarXiv:2610.29997
The authors present DONIC, a single generalist policy for bimanual dexterous reference tracking, trained jointly across operation categories and human-motion sources. They also introduce BiHOI, a multi-view dataset of 1,575 bimanual interaction sequences over 229 objects, with reconstructed interactions retargeted to robot references.
124 upvotes
AIarXiv:2610.29989
The authors propose Future Chunk Prediction (FCP), an auxiliary training objective that is optimized jointly with teacher-forced denoising in autoregressive video diffusion models. FCP predicts temporally separated future chunks without access to the intervening visual context, using supervision taken directly from training videos. Autoregressive inference is left unchanged, and no extra pretrained visual encoder is needed.
108 upvotes
AIarXiv:2610.27851
The authors show that JEPA-style latent world models can learn across diverse domains and data sources by adopting the language-conditioned interface of video generation models. They introduce LaLa-World, which describes both numerical actions and visual changes in language and builds the future predictor on a language model, so one predictor trains jointly on heterogeneous datasets. The pretrained model is also used to initialize LaLa-Policy, which jointly predicts actions and future visual representations.
27 upvotes
AIarXiv:2610.29722
The authors built LASAGNA, an open-source framework for controlled RAG evaluation, and tested 450 pipeline configurations across eight benchmarks and corpora of 5K to 200K documents. They examined how embedding, chunking, retrieval, reranking, query augmentation, context compression, and agentic verification behave when combined rather than tested in isolation.
26 upvotes
AIarXiv:2610.30320
The authors introduce FRED, a framework that trains World Action Models to improve their future video representations extracted from a single denoising step. FRED uses self-distillation, predicting future representations from noise with targets taken from ground-truth future observations, and conditions the video model on historical observations for temporal context.
23 upvotes
AIarXiv:2610.30113
The authors built a benchmark and public leaderboard for end-to-end quantitative research on real A-share market data. Models write features and stock-scoring code using feedback only from 2022–2023, then freeze one strategy that is evaluated on the sealed 2024 year with a cash-account evaluator that applies trading constraints, costs, and order-book-based market impact.
21 upvotes
AIarXiv:2610.29545
The authors propose PACE, a post-hoc four-bit quantizer for the recurrent states of linear-attention language models. PACE rounds the state only every few generated tokens, buffering the intermediate tokens, and shares one quantization scale per column because large gated-delta state entries align in columns; heads that retain error longest are kept at eight bits.
19 upvotes
AIarXiv:2610.28402
The authors introduce StreamEval, a unified benchmark for interactive streaming video generation. It contains 300 scripted test cases across five tasks (avatar dialogue, avatar performance, world generation, instruction editing, reference editing), scored on ten task-specific metrics across four dimensions.
330 upvotes
AIarXiv:2610.28577
The authors propose cache engineering, which lets LLM agents directly manage their active KV cache by dropping selected cache entries from attention and repositioning retained keys, without editing the trajectory or re-prefilling. They build the AKV serving engine, which exposes a stateless API while handling stateful cache reuse, eviction, and restoration internally.
111 upvotes
AIarXiv:2610.28698
The authors run a controlled study of continued pre-training (CPT) for embodied game agents, varying data volume, data diversity, and model size across 120 configurations. They evaluate with validation losses at three levels and with gameplay benchmarks, from worlds seen in training to unseen ones, and fit an empirical scaling law.
46 upvotes
AIarXiv:2610.28889
The authors propose Policy-Mask, a budget-conditioned policy trained on the downstream objective that jointly selects a timestep, a network block, and a channel subset from a frozen diffusion model under a fixed channel budget. They normalize the channel-score term by subset size so that the three policy heads update on comparable scales. The method is also adapted to diffusion transformers such as PixArt-Σ and SD3.5 MMDiT.
43 upvotes
AIarXiv:2610.27598
The authors compare Local Experts (layer-specific, as in mainstream MoE models) against proposed Meta Experts that are shared across layers. They run a multi-dimensional empirical study and add a theoretical analysis of residual pathways.
25 upvotes
AIarXiv:2610.28416
The authors build Code2Video-Compass, a benchmark of 2,104 tasks for generating videos by writing executable code across text-to-video, image-to-video, and video-to-video settings. Tasks span 10 categories and 24 subtypes, and are scored with 12 objective metrics and 5 subjective metrics, the latter judged by an MLLM.
24 upvotes
RoboticsarXiv:2610.28589
The authors introduce AgenticRoboRL, a framework for post-training robot planners across different robot harnesses. It combines supervised fine-tuning with multi-turn reinforcement learning, using harness-specific adapters so one RL pipeline can work with both VLA skill invocation and control-code generation while execution components stay fixed.
20 upvotes
AIarXiv:2610.30181
The authors introduce OpenDS, a task for evaluating multi-step deep search on the live, open Web, where agents must find, access, and integrate evidence from multiple independent sources to reach a unique factual answer. They build OpenDS-World, a 200-task benchmark spanning diverse modalities, file formats, and interactive access requirements, human-verified for evidence validity and answer uniqueness, and evaluate 12 frontier models under a unified tool setup.
149 upvotes
AIarXiv:2610.26659
The authors introduce SpatialScout, a benchmark of 80 indoor scenes with 3,140 target objects. VLM agents control a camera to explore each scene and estimate object centers in a shared coordinate frame under bounded interaction budgets.
129 upvotes
AIarXiv:2610.26828
The authors present WildAWM, a scalable pipeline that synthesizes controllable, executable computer-use environments and tasks spanning GUI, MCP service, and command-line interfaces. The pipeline is powered entirely by open-source models and adds execution-based quality gates for solvability, difficulty, and verifier reliability.
64 upvotes
AIarXiv:2610.26233
The authors propose Heartbeat Thinking, a method for voice agents that starts reasoning and tool calls while the user is still speaking. Instead of cutting off generation at each speech update, periodic Thinker invocations read the latest partial transcript and a shared record of completed reasoning and tool results, while earlier invocations and dispatched tools are allowed to finish.
18 upvotes
AIarXiv:2610.19705
The authors propose RouteWeaver, a single LLM routing policy that combines single-round, multi-round, and agentic routing modes and learns both which mode to use and which models to pick within it. They train it with COMET-GRPO, which splits the trajectory-level advantage into separate signals for mode selection and within-mode execution, plus a progressive exploration schedule that moves from externally assigned modes to autonomous choice.
17 upvotes
AIarXiv:2610.25100
The authors propose Deep Fusion Memory (DFM), a parametric memory architecture that trains a causal memory predictor to generate memory from preceding context, supervised by retrieved chunk embeddings. The predicted memory is shared across each chunk's tokens and fused into hidden states via gated cross-attention, with the backbone frozen and no retrieval at inference.
39 upvotes
RoboticsarXiv:2610.23479
The authors test how far a vision-language-model coding agent can go when it controls a two-arm robot through a command line instead of a computer. The agent reads three calibrated RGB cameras, and after each failed attempt it writes or fixes small tools that use triangulation and plane intersection to locate objects. The tools are built on ten layouts per task, frozen, and then tested on ten new layouts.
84 upvotes
AIarXiv:2610.23121
The authors introduce Scene-Designer, an agentic framework that learns both to design 3D indoor scenes and to critique them, using two new datasets of design and critique experiences. They train it with Scene-GRPO, a Critic-guided variant of GRPO, so it makes better design decisions and needs fewer trial-and-error corrections.
40 upvotes
AIarXiv:2610.23181
The authors propose TreeOPSD, a critic-free framework that branches alternative continuations from shared prefixes, backs up terminal rewards into turn-level credit, and maps each credit to a bounded, closed-form teacher distribution. Minimizing the divergence between this teacher and the student recovers the policy-gradient direction while keeping sequential mini-batch updates near a finite target.
27 upvotes
AIarXiv:2610.23542
The authors introduce MIST, a memory test that separates whether a scene was set, whether the requested change was carried out, and whether the final frame shows the result. They then propose M3-Agent, which recursively revises the details of the instruction so the frozen model carries out the requested change.
19 upvotes
AIarXiv:2610.23230
The authors build RAP, a dataset of over 25K verified paper-poster pairs from machine learning conferences (2022-2026), where each human poster is converted into styled HTML/CSS with refillable slots. They also release a 150-paper held-out benchmark and use the corpus for two approaches: PosterRAG, a training-free multi-agent pipeline that retrieves and refills exemplars, and an SFT+RL model.
19 upvotes
AIarXiv:2610.23787
The authors propose DLite, a speculative decoding framework for agentic AI that reuses hidden states from the target model and makes local predictions without maintaining the growing historical context. A lightweight sequential head restores causal dependencies within each block, and a tailored two-stage training strategy trains this architecture.
18 upvotes
AIarXiv:2610.23194
The authors introduce OfficeTown, a framework that generates simulated professional organizations from a natural-language description of their purpose and activities. Members coordinate work, delegate bounded assignments to assistant agents, and review deliverables, while artifacts, decisions, and open matters persist and shape later work. The team converts these work requests, context, and review feedback into OfficeTown Bench and training data.
16 upvotes