The paper introduces PACE, a Pareto‑adaptive compression framework designed for native multimodal large language models (MLLMs) that directly tokenize pixel patches without external vision encoders. By analyzing token discriminability and heterogeneous redundancy across spatial dimensions, PACE decomposes attention into height‑ and width‑specific saliency maps and applies a Pareto‑optimization strategy to allocate token budgets, achieving near‑lossless performance with only 25‑50% of visual tokens and more than 2× FLOP savings. Experiments demonstrate that PACE not only preserves but sometimes improves benchmark results compared to existing ViT‑based compression methods.
By Lei Tian, Chunzheng Zhu, Kai Wu, Yifan Zhang, Haihua Yang
Web agents are increasingly deployed to complete tasks autonomously on real websites, where adversarially injected content can induce them to perform unauthorized actions. Existing evaluations primari...
Estimating whether a problem is too complex for a large language model (LLM) before the model attempts it can guide model routing, how to allocate test-time compute, and suggest the need for new tools...
An LLM agent that helps a patient get care must do two things. It must reach the correct outcome, and it must follow the necessary procedure. For example, the agent must get the patient's permission b...
The paper introduces the Controlled Attribution Benchmark (CATT), a synthetic evaluation framework that fine‑tunes language models on entirely fictitious knowledge to provide clear ground truth for data attribution. By systematically varying single dataset properties—such as document redundancy and lexical distractors—CATT reveals that these factors consistently degrade attribution performance. The study also explores two underexamined design choices: using negative log‑odds gradients instead of cross‑entropy gradients, and aggregating only the top token attribution scores, which together improve ground‑truth recall by 35–72 percentage points and enhance counterfactual retraining performance on FTRACE‑TREX by up to 4.9×.
The paper introduces world2vec, a method for learning mechanistic world models that capture true system dynamics without relying on reward signals. By learning latent dynamics from observations and actions, world2vec achieves accurate long‑horizon predictions with significantly fewer parameters than existing models such as DreamerV3, TD-MPC2, and LeWM. This efficiency enables state‑of‑the‑art control of continuous environments from pixel inputs and precise system identification across 71 benchmark attractors, illustrating the potential of world models for scientific discovery.
The Ramanujan Challenge benchmark tests AI on research‑level number theory problems, including unpublished results and open conjectures. Analysis of submissions shows that AI models, CAS, and formal tools can produce correct solutions that differ conceptually from official ones, even proving a published open conjecture. To reliably judge these complex proofs at scale, the authors propose an AI‑assisted evaluation pipeline that reduces false positives in natural language proofs from 23% to 6% and improves code judgment accuracy from 46% to 58%.
The paper introduces Rewindable Interactive Proofs (RIPs), a new model for scalable oversight where a weak verifier can rewind a powerful AI prover to an earlier state during interaction. RIPs enable efficient verification of computations whose outputs are robust to small changes in oracle answers, achieving near-linear verifier time without relying on cryptography. The authors implement the protocol, demonstrate a preliminary MNIST experiment, and show that an RIP can outperform a state‑of‑the‑art deployed proof system.
BURGER is a language‑conditioned robotic drawing benchmark that uses a prompt‑driven pipeline to generate continuous drawing trajectories in both simulation and the real world. The benchmark provides 28,200 simulated demonstrations across 282 tasks, allowing evaluation of vision‑language‑action models on data scale, language grounding, and multimodal behavior. Experiments show that some models improve with more diverse data, capture complex drawing behaviors, and generalize to unseen compositions, with a trained policy successfully deployed on a Franka Panda arm.
LucidLyra is a framework that reconstructs 3D scenes from video latents in a feed‑forward manner while simultaneously providing open‑vocabulary semantics and instance cues. By distilling knowledge from a camera‑controlled video diffusion model and frozen 2D vision foundation models, it jointly infers 3D Gaussians that encode geometry, appearance, and semantic information without requiring per‑scene optimization or manual 3D annotations. The resulting scenes support novel‑view rendering, open‑vocabulary segmentation, and object grouping, and can be extended to monocular video for dynamic 3D reconstruction.
The paper introduces neural triage decoding, a new framework that uses syndrome consistency to adaptively allocate neural computation for high-rate quantum LDPC codes. By combining a local neural message-passing network with iterative refinement and a rare diversification stage, the Triage decoder achieves state‑of‑the‑art accuracy while reducing decoding latency by one to two orders of magnitude. The approach is validated on ultra‑high‑rate qLDPC codes such as lifted‑product and Kasai codes, and demonstrates robustness to mismatched error priors.
The paper introduces Continual Orthogonal Projection (COP), a gradient‑projection method that limits each unlearning update to the orthogonal complement of previously cached activation subspaces, thereby preventing interference between successive deletions. COP requires only a low‑rank basis per layer, avoiding the need for past gradients or previous forget sets, and can be applied to any differentiable‑loss unlearning method without altering its objective. Experiments on TOFU and RWKU benchmarks demonstrate that COP maintains model utility across sequential unlearning requests while preserving forgetting effectiveness and preventing collapse of standard objectives.
The paper introduces KITScenes Aerial, an extension of the KITScenes Multimodal dataset that adds synchronized aerial imagery and high‑resolution neural representations to existing onboard sensor data, vehicle poses, and Lanelet2 map annotations. It evaluates mapping across all 53 KITScenes map classes, showing that aerial context significantly improves fine‑grained and extended‑range mapping, with higher‑resolution imagery boosting performance further, especially for challenging elements like traffic lights. The study also assesses cross‑view localization under geographically disjoint conditions, achieving a mean translation error of 5.13 m with a fine‑grained feature‑matching baseline.
HiFOR introduces a hierarchical foresight framework for Vision‑Language‑Action models, combining short‑term state evolution, semantic milestones, and final goal predictions via a Hierarchical Dependency Mask. It structures future latent representations across horizon, view, and spatial dimensions, enabling fine‑grained retrieval. An action‑guided factorized retrieval mechanism then selectively pulls relevant future information during denoising to guide action generation, improving robotic manipulation performance across simulation and real‑world tasks.
The paper introduces EpiCoach, an agentic system that treats a single athlete as the subject of accountable N‑of‑1 science. It uses a mechanistic digital twin built from wearable data to make probabilistic predictions about future training sessions, and an evidence‑gated hypothesis registry records each claim and its outcomes. Over eighteen months, the system’s predictions were evaluated both retrospectively and prospectively, showing competitive performance and demonstrating a framework for auditable, falsifiable personalization in digital health.
The paper introduces a unified Hessian-guided framework for post-training quantization of large language models. It derives a Hessian-based approximation of output changes due to joint weight and activation quantization, enabling simultaneous optimization of quantization ranges, linear transformations, and weight rounding under a single output-sensitivity objective. The method remains practical for modern models and consistently outperforms existing transformation and rounding techniques in both channel-wise and block-wise settings.
ScienceJudge introduces Pattern Perturbation Augmentation (PPA) to generate hard negative examples for sentence‑level scientific claim verification by modeling reporting errors as faithfulness‑breaking transformations. PPA applies these patterns to supported claims, creating controlled hard negatives and expanding coverage through pattern‑level, operation‑level, and cross‑domain augmentation. The resulting ScienceJudge system, which ensembles PPA‑trained vision‑language predictors with instruction‑guided LLM predictors, achieves state‑of‑the‑art performance, improving the average score from 0.8124 to 0.8692 and ranking first on the NLPCC 2026 Task 10 Track 1 leaderboard.
The paper introduces Coupled Guidance, a flow‑matching‑specific guidance mechanism that addresses instability issues arising from directly adapting diffusion guidance methods. By leveraging the coupling structure inherent to flow‑matching training, the authors provide a theoretical analysis showing that naive diffusion guidance breaks the training distribution and weakens control. Extensive experiments on robotic collision avoidance, dexterous manipulation, and image generation tasks demonstrate that Coupled Guidance improves sample quality and guidance stability without extra computational overhead.
The paper reports the submissions of team nucli to the HECKTOR 2026 challenge, which evaluates machine learning methods for head‑and‑neck cancer imaging. The team used self‑supervised pre‑trained encoders, training a 2D DINOv2 encoder on PET scans and combining a DynUNet with a SwinUNETR for segmentation, achieving a mean Dice of 0.5409. For TN staging, they fused geometric features from segmentation masks with PET‑DINOv2‑MIPs deep features via logistic regression, obtaining balanced accuracies of 0.3962 (T‑stage) and 0.6432 (N‑stage). For recurrence‑free survival prognosis, they fine‑tuned the PET‑DINOv2‑MIPs ViT‑S encoder with an MLP head under a CoxPH loss, achieving a Harrell's C‑index of 0.6577 and a weighted challenge score of 0.5802.
The paper introduces CoC-CFG, a method that applies classifier‑free guidance to the reasoning trace of vision‑language‑action (VLA) policies, weighting each Euler step to modulate the influence of reasoning. By treating the weight selection as a Markov decision process with measured action values, the authors develop Q‑adaptive guidance that conservatively improves the controller. Experiments on the Alpasim driving simulator and the Waymo end‑to‑end benchmark show that Q‑adaptive guidance outperforms fixed weights and online RL baselines, achieving a new state‑of‑the‑art score on the hidden test set.