AI Papers · Daily Picks

ATAS improves long-form video consistency by compiling screenplays into referenced shot requests, retrieving spatial and temporal evidence from scene and event memories, and conditionally updating memory after generation.

2610.20571 · ▲ 323 · Code

GroundingPI is a 4B grounding foundation model that improves precise object localization and downstream robotic and autonomous driving performance through multimodal pretraining and reinforcement learning.

2609.39601 · ▲ 314 · Code

Multimodal language models use an intrinsic linear subspace called S-Space to encode object coordinates, supporting spatial perception, reasoning, and response generation, while also revealing bottlenecks in viewpoint transformation.

2610.13527 · ▲ 294 · Code

Jeva is a decision-native architecture that jointly enforces null-awareness, permutation invariance, option scalability, and cross-request caching for structured decision-making.

2610.18773 · ▲ 276 · Code

Player-1 improves game control by combining supervised fine-tuning on gameplay trajectories with reinforcement learning across synthetic environments, achieving substantial gains over demonstration-only baselines.

2610.20584 · ▲ 270 · Code

Researchers propose multi-level metrics to evaluate LLM no-thinking behavior, finding that explicit reasoning persists despite controls and varies with task openness, establishing no-thinking as a distinct capability requiring systematic assessment.

2610.11765 · ▲ 244 · Code

Road MOPD reduces gradient conflicts in multi-teacher distillation by using cross-teacher agreement for token-level credit assignment and auxiliary gradients to constrain conflicting updates.

2610.13521 · ▲ 233

AgentSWE evaluates whether coding agents build software that truly meets user requirements across creation, editing, and optimization stages, revealing widespread failures to deliver faithful behavior.

2610.13361 · ▲ 227 · Code

TerminalHorizon automates long-horizon terminal agent data synthesis by reconstructing environments around omitted work and progressively expanding tasks, yielding substantial gains across coding, production, and reasoning benchmarks.

2610.13166 · ▲ 160

DiffusionART improves reinforcement learning fine-tuning of diffusion models by removing hidden consistency forces and enabling stable, efficient training without EMA tuning or differentiable rewards.

2610.12572 · ▲ 148 · Code

VideoJev converts video observations into structured decisions by scoring temporal intervals and text options, enabling efficient reusable inference for visual agents.

2610.18615 · ▲ 145 · Code

Urbex reframes city-scale 3D grounding as an active search process using a vision-language agent with reinforcement learning, improving accuracy and speed without precomputed proposals.

2610.18356 · ▲ 116 · Code

SoftwareGym automates synthesis of diverse, executable GUI applications to provide scalable training data that improves computer-use agent performance on out-of-domain benchmarks.

2610.12611 · ▲ 107 · Code

GroundAnything is a 4B-parameter visual grounding model that uses bidirectional diffusion and blockwise denoising to enable fast parallel decoding with precise localization, outperforming autoregressive counterparts across 30 benchmarks.

2609.39600 · ▲ 94 · Code

Doc2Harness is a deterministic two-stage benchmark that evaluates long-horizon harness engineering by having models build and run harnesses across isolated containers, revealing substantial performance gaps among frontier models.

2610.20344 · ▲ 93 · Code

Harness² improves agent execution harnesses through recursive task-level refinement and continual harness-level learning, boosting performance across benchmarks without ground-truth verifiers.

2610.14224 · ▲ 90 · Code

Reflex is a three-stage framework that enables humanoid robots to catch thrown boxes using egocentric vision by decomposing whole-body control, dynamics inference, and visual perception into sequentially trained stages.

2610.12772 · ▲ 88 · Code

GreenWAM is a unified world foundation model for robotics that uses a shared anatomical action space and joint training to serve as both a predictive world model and a control policy across diverse robot bodies.

2610.18774 · ▲ 88 · Code

Tracking-Value PPO improves reinforcement learning for reasoning by modeling value evolution through cumulative log-odds updates, yielding better accuracy and generalization.

2610.16794 · ▲ 78 · Code

EmbodiedCLBench evaluates continual learning in embodied agents by testing compositional skill reuse across basic and complex tasks, revealing stable but non-scaling experience benefits.

2610.13021 · ▲ 73 · Code

Researchers benchmarked ten self-distillation methods against GRPO across multiple agent tasks and model sizes, finding limited consistent gains, hidden training instability, and no reliable cost-to-performance benefit.

2610.19789 · ▲ 72 · Code

DoSA accelerates sparse attention by using a certificate to conditionally score only eligible tokens, preserving exact selection while reducing computation.

2610.13990 · ▲ 71 · Code

CriticOPD improves on-policy distillation by using teacher critiques to restore recoverability, enabling students to revise incorrect reasoning and achieve stronger mathematical reasoning performance.

2610.20572 · ▲ 71 · Code

Vela converts real-world user behavior records into training tasks and trajectories, improving large language model performance on user-related and general benchmarks through supervised fine-tuning and reinforcement learning.

2610.14709 · ▲ 71 · Code

VisionWeave introduces elastic visual representation weaving to adaptively allocate visual tokens by granularity, reducing token usage by 43% while preserving nearly all model performance across diverse tasks and serving conditions.

2610.07987 · ▲ 68 · Code

LeWAM is an end-to-end joint-embedding predictive architecture that jointly learns latent dynamics and action flow matching for efficient planning and direct control.

2610.12818 · ▲ 65 · Code

StereoBind binds visual motion to stereo sound generation via motion tracks and specialized encoders, improving spatial alignment in joint video-audio synthesis.

2609.38748 · ▲ 65 · Code

LatentLoop improves looped Mixture-of-Experts by compressing the first loop visit, reallocating saved computation to larger hidden dimensions, and using separate attention parameters, yielding lower validation loss and better benchmark scores.

2610.19975 · ▲ 64 · Code

Cortex consolidates context-induced improvements into mergeable low-rank or structural updates, preserving non-target capabilities across sequential edits and diverse model scales.

2610.20399 · ▲ 63

Multics introduces hierarchical group-level communication and dynamic grouping to scale multi-player world models efficiently while reducing interference and maintaining consistency.

2610.17201 · ▲ 57 · Code

ProgramGym introduces scalable verifiable training environments and a role-isolated agent workflow to improve long-horizon program reproduction from scratch.

2610.13395 · ▲ 54 · Code

ActionOPD aligns smaller language-model agents with teachers via semantic action-level on-policy distillation, improving out-of-distribution tool-use performance without requiring shared tokenizers or teacher probabilities.

2610.19820 · ▲ 53 · Code

Work2Skill captures human GUI demonstrations to induce reusable agent skills, improving execution performance on desktop tasks through workflow rules and step dependencies.

2610.12984 · ▲ 51 · Code

Agent-Reflex trains interactive agents to reflect and recover from noisy tool feedback and user corrections through skill discovery, supervised fine-tuning, and on-policy optimization.

2610.17953 · ▲ 49 · Code

AReA improves high-resolution image generation with diffusion transformers by adapting attention from global to local across denoising steps without extra training.

2610.17227 · ▲ 49 · Code

Multi-Stream Attention Residuals extend depth-wise attention to multi-head retrieval with adaptive mixing, improving efficiency and performance in large-scale pre-training.

2610.13630 · ▲ 47 · Code

RSI-Anything leverages predictive world models to guide action selection and adaptive learning, enabling recursive self-improvement across coding, interface navigation, laboratory planning, and scientific optimization tasks.

2610.13912 · ▲ 46 · Code

This work proposes a two-axis framework for multimodal agent memory covering information formation and adaptive management to support lifelong learning.

2610.19366 · ▲ 45 · Code

LensErase is a mask-based erasure foundation model pretrained on a large paired dataset and refined with reinforcement learning to handle diverse real-world removal tasks while preserving surrounding content.

2610.18072 · ▲ 36 · Code

PolyOCR is a family of unified OCR foundation models that integrates instruction following, large-scale data synthesis, and competence-guided policy optimization to achieve strong recognition, parsing, and reasoning across diverse visual text tasks.

2609.37712 · ▲ 36 · Code

This survey unifies research on agentic robots by tracing the evolution from robot policies to integrated systems using vision-language-action models, world models, and structured functional domains.

2610.20442 · ▲ 29 · Code

Recent advances in robot learning have driven rapid progress in humanoid control, enabling increasingly agile behaviors in the real world. However, humanoid control networks are typically obtained through conventional training procedures, while comparatively little attention has been given to directly modeling the distribution of their learned weights and generating functional networks in weight space. In this work, we investigate whether high-performing humanoid control networks can be generatively modeled directly through their weights. We adopt a latent diffusion framework that couples a linear autoencoder with a Diffusion Transformer. The autoencoder compresses actor network parameters into a compact latent space, and the Diffusion Transformer models the distribution of these latents. Trained on networks collected from reinforcement learning for the Unitree G1 robot, our approach generates deployable networks that perform various motions in the real world, including walking, dancing, and spin kick. Beyond generation, we further study the learned weight space to better understand the structure captured by these humanoid control networks. We identify a structural principle in the first-layer weights that consistently reflects the robot's kinematic prior. Notably, this kinematic prior emerges under both Reinforcement Learning (RL) and Imitation Learning (IL) training, suggesting that it is not specific to a particular optimization process but may instead reflect common knowledge associated with humanoid embodiment. Together, these results demonstrate the feasibility of generative modeling in humanoid network weight space, while providing initial insights into the information encoded by their learned weights. Additional results are available on our project website: https://humanoid-weight-diffusion.github.io/

2610.20606v1

Mobile Manipulation (MoMa) is a cornerstone of autonomous robotic systems, yet real-world fine-tuning remains a formidable challenge because of high-dimensional action spaces and intricate coupling between heterogeneous modules. In this paper, we propose HiRe-MoMa (Hierarchical Residual Mobile Manipulation), a systematic framework for stable and sample-efficient real-world fine-tuning of complex MoMa systems. At its core, HiRe-MoMa adopts a Hierarchical Residual Architecture that fine-tunes a frozen base policy. A Policy-Agnostic Residual learns additive corrections atop the base policy, while Hierarchical Decoupling splits the residual into mobile-base and manipulator-arm heads with unidirectional conditioning, mitigating gradient interference and aligning corrections with the platform's physical hierarchy. To improve sample efficiency in data-expensive real-world settings, we propose Expert-Guided Value Alignment, which leverages sparse human interventions with under 20% training coverage to reshape the critic's value landscape, effectively guiding autonomous exploration through expert priors. Extensive real-world experiments on the TIAGo++ platform demonstrate that HiRe-MoMa consistently enhances task performance with only around 30 minutes of online interaction, regardless of the base policy architecture. To our knowledge, this is among the first frameworks to enable efficient on-robot fine-tuning for mobile manipulation in the real world, providing a practical paradigm for deploying complex robotic systems in dynamic, human-centric environments. Further details can be found on our project website: https://hire-moma.github.io/

2610.20607v1

Training language agents is bottlenecked by the high cost of long-horizon rollouts, yet standard agentic reinforcement learning (RL) learns only from a terminal scalar reward, leaving the supervision in intermediate steps largely unused. Prior work exploits this supervision through value targets that estimate per-step success probabilities or observation targets that fit environment response. Both are hard to construct well: (1) value targets come from the same single scalar signal for each step in the trajectory, and (2) observation targets, such as web pages and terminal logs, are long and noisy and reach only one step ahead. We propose \name, which adds a third kind of multi-step action targets. On successful on-policy trajectories, \name trains the policy backbone to generate the action sequences that follows each history. Each action is also supervised when forecast from the preceding histories, forming a temporal ensemble that the shared backbone absorbs during training, and this enables acting and forecasting to co-evolve. The policy still acts one step at a time at inference, with no overhead. Across five benchmarks (ALFWorld, ScienceWorld, WebShop, $τ^2$-bench, and AppWorld), \name (1) achieves stronger aggregate performance than GRPO and RL baselines with predictive auxiliary targets across simulated environments, while improving overall performance on both realistic applications; (2) delivers up to $1.9\times$ and $2.0\times$ higher sample efficiency than ECHO and GRPO, respectively, on ScienceWorld; and (3) achieves these gains with a single policy backbone, without additional models, separate pretraining, or inference overhead. The codebase is available at https://github.com/ulab-uiuc/CoPE.

2610.20597v1

Scaling up transformers has produced models with broad capabilities, deployed across heterogeneous hardware and fluctuating workloads. Yet, model families offer only a few fixed sizes (e.g., S, B, L), forcing users to commit to a model based on hardware and latency constraints. This rigidity is often suboptimal under varying query loads. We introduce a retraining-free method that elastifies pre-trained transformers into a single model that adapts on-the-fly across a continuum of computational budgets. Our key contribution is a differentiable relaxation of the functional unit-ranking problem: unlike prior work, which optimizes a single importance weight per MLP block or attention head, we rank every individual MLP neuron and attention head jointly, on a common scale, by minimizing the model's loss under soft pruning across many sparsity targets, all at once. Across 27 vision and text transformer encoders ranging from 22M to 8B parameters, our method yields smooth degradation up to 60% sparsity and outperforms one-shot baselines under deep pruning. For example, pruning DINOv3 ViT-H+/16 to 50% sparsity removes 427M parameters while reducing ImageNet-1k linear-probing accuracy by only 0.8%, without the need for weight correction or retraining. At 60% sparsity, a pruned AugReg ViT-B/16 model achieves up to $1.80\times$ higher throughput and $1.26\times$ lower single-query latency on an A100 GPU. Elastic models are produced in minutes, with construction time scaling sub-linearly with parameter count.

2610.20588v1

Long contexts enlarge key--value (KV) caches, increasing memory use and attention costs during generation. Compressing these caches must preserve information for unseen questions. We introduce **Latent Compaction**, which learns a compact document-specific cache through test-time training. We initialize cache entries using reconstruction importance, teacher attention, and key-space diversity. Training first fits attention outputs and log-normalizers to those of the full cache, with reconstruction-selected entries held fixed. It then jointly optimizes the entire cache through end-to-end distillation from the full-cache model. Model weights remain frozen throughout, and training uses only generic prompts about the document, without access to downstream questions or answers. Experiments on RULER, LongBench, and Phonebook Lookup demonstrate substantial gains over budget-matched compression baselines. On a RULER, for example, our method achieves 96.7% accuracy at 32$\times$ compression, compared with 40.0% for KVZip+. The method also transfers across model scales and families, with retrieval gains on both Qwen3 and MiniCPM5. These results show that treating the KV cache as a neural network and optimizing it with test-time training can preserve retrieval accuracy under aggressive compression. The code is released at https://www.github.com/THUNLP/latent-compaction

2610.20582v1

Brazilian Portuguese remains underserved by open-source language models, and the few that are safety-aligned rarely document what did not work or validate comparisons with statistical significance. We present Manacá-1B-Instruct-2, the second-generation post-training alignment of Manacá-1B (1.72B parameters, trained from scratch for PT-BR), with an SFT corpus of 291,223 examples. We document three DPO attempts, all negative, including a case where the training reward margin reaches 1.0 while free generation does not change at all, and show that safety came from a dedicated safety-SFT (LLM-jp AnswerCarefully style), refined over three iterations until closing the safety-helpfulness frontier. A controlled scan of the fraction of general data in the safety-SFT mix (10% to 60%) reveals a non-monotonic over-refusal curve with a knee point at 40-50%. The final model is compared against four external PT-BR models on the same items, with Wilson confidence intervals and paired McNemar tests with Holm-Bonferroni correction: the lower bound of its safety interval (94.1%) lies above the upper bound of all four baselines. We also report a methodological finding for LLM-judge evaluation: on two of six safety items in MT-Bench-PT, the judge categorically refuses to assign a score regardless of the evaluated response, silently reducing the effective N of that category in every prior report of the project. We complement this with an n-gram decontamination check and a Chinchilla-anchored compute-budget table that attributes the gap in general capability to pre-training scale rather than alignment methodology. All code, raw data, logs, and scripts that produce every reported number are publicly released.

2610.20570v1

Speculative decoding accelerates autoregressive generation by drafting multiple tokens per step, yet feature-based drafters rely on target hidden features whose contextual emphasis may not align with the needs of subsequent draft positions. Hybrid linear attention (HLA) models maintain compact recurrent states that accumulate historical associations, offering an efficient source of complementary context for speculative drafting. We introduce ReaDy, the first speculative decoding method, to the best of our knowledge, to leverage HLA recurrent states for position-specific context retrieval during drafting. ReaDy equips each draft position with a learned query to retrieve supplementary context directly from the target's committed recurrent states. A lightweight state reader integrates this context into draft representations through residual updates, while low-rank state reads and fused operators keep additional computation and I/O overhead low. Building on DSpark's parallel backbone and sequential prediction head, ReaDy employs a lightweight two-stage adaptation strategy, with reader capacity, state sources, and residual placement examined through ablation studies. Across seven benchmarks spanning mathematics, code generation, and dialogue, ReaDy achieves an average $3.16\times$ decoding speedup over autoregressive decoding with Qwen3.8-27B on a single NVIDIA B200 GPU. It outperforms DSpark on every benchmark, improving throughput by 7.97% and acceptance length by 8.50% on average. Our code is available at https://github.com/StartLuxLabs/ReaDy.

2610.20503v1

Pretrained vision-language-action models (VLAs) often fail midway through unseen long-horizon tasks. Recursive self-improvement (RSI) offers a remedy, but embodied RSI methods either refine the harness around a frozen executor and so cannot add a missing motion, or train the executor on persistent failures without asking whether a failure lies in how the executor is called or in what it can do. We ask when a self-improving embodied system should change the weights of its executor, and propose AdaRSI, an on-demand RSI framework for embodied agents. A coding agent first improves its skills and harness, then attributes persistent failures by retrying failing sub-steps from saved states, and fine-tunes only the sub-steps that remain unreliable, with demonstrations it collects itself and experience carried over from earlier tasks. On all 50 target tasks of RoboCasa365, with GR00T N1.5 as the executor and no human demonstrations, AdaRSI improves the VLA on all three suites and raises its success on unseen composite tasks from 4.6% to 13.8%. Compared with fine-tuning every sub-step, it reaches slightly higher success while using only 28% of the GPU time and 22% of the self-collected demonstrations, and fine-tuning without a harness trains more yet gains far less.

2610.20474v1

Generating novel views from limited observations requires preserving both the appearance and spatial structure of a scene. Although recent approaches exploit fine-grained geometric cues, they do not explicitly model high-level semantic and spatial properties such as object identity, attributes, and inter-object relationships. In this work, we investigate whether 3D language understanding can provide complementary supervision for novel-view generation. We introduce Ask4NV, built on a pretrained unified multimodal model with a Mixture-of-Transformers (MoT) architecture. Our model incorporates camera geometry into both VAE and ViT representations and leverages understanding features during generation through joint self-attention with a modality-aware attention masking strategy. Experiments demonstrate strong image quality and geometric consistency in novel-view generation, alongside competitive 3D language understanding. Further analysis shows that joint training improves cross-view correspondence in VAE representations and similarity in a 3D-aware feature space.

2610.20610v1