Leaderboards, eval harnesses and ablations — the contested business of deciding which model is actually better.
Out-of-distribution detection in continual learning (OOD-CL) is highly challenging, as it requires (i) learning a sequence of tasks without catastrophic forgetting, and (ii) identifying unknown sample...
Histopathology is central to cancer diagnosis, but morphology in Hematoxylin and Eosin (HE)-stained images does not directly reveal the molecular states that characterize disease heterogeneity and inf...
Language models increasingly use skills that provide reusable instructions and procedures for task execution. After a relevant skill has been selected, the model may need to apply its governing proced...
The paper introduces an automatic metric for evaluating chess explanations based on chess-engine evaluations and finds that current language models perform poorly, relying on surface-level observations. It presents BinProp, a training-free method that helps models discover reusable chess concepts, significantly improving explainability across multiple models. BinProp also generalizes to other board games such as Go, draughts, and shogi without extra tuning.
Xiaomi‑Lens‑0 is a 4B multimodal large language model built on Qwen3.5 that predicts region masks directly from its visual representations, eliminating the need for a separate segmentation encoder. By combining vision‑encoder patch features with visual latents and mask‑token states, the model interleaves reasoning with region‑specific latents, enabling grounded chain‑of‑thought reasoning. Across all RefCOCO, RefCOCO+, and RefCOCOg splits, Xiaomi‑Lens‑0 outperforms Qwen3.5 by an average of 5.6 points in referring comprehension and achieves the highest segmentation cIoU among compared models, while also surpassing Qwen3.5 on all perception benchmarks.
HyperPersona is a multi‑level hypergraph framework that models the hierarchical structure of text—document, sentence, and word—using hyperedges for documents and sentences and nodes for words. It employs a transformer‑based graph encoder to capture interactions across these linguistic layers, producing context‑sensitive, structurally grounded features for automatic personality prediction. Experiments on the Big Five personality dimensions demonstrate that HyperPersona, relying solely on text, outperforms state‑of‑the‑art baselines by effectively integrating multi‑level linguistic cues.
MGE-PO introduces an entropy-aware credit assignment framework for mathematical reasoning, using a four-quadrant advantage shaping that blends verified correctness with rollout generation entropy. The method assigns larger bonuses to low-entropy successful responses and larger penalties to high-entropy failures, guided by an online entropy-based question curriculum. Experiments on Qwen3-8B-Base and Qwen2.5-7B-Base show significant macro average and pass rate improvements across multiple benchmarks.
The paper introduces NoLBERT, a language model pretrained only on financial news from 1976 to 1995, and adapts it into a sentence encoder via contrastive fine‑tuning. It compares NoLBERT’s topic extraction to LDA, MiniLM, and FinBERT on 1.26 million articles from 2012 to 2024, converting topics into attention shocks and stock exposures to forecast returns of the 1,000 largest US stocks across multiple models and rebalancing frequencies. While NoLBERT produces the most stable topics and performs comparably to LDA, the best‑validated model still loses money, and an equal‑weight blend of all readers yields a monthly net Sharpe ratio of 1.12, significant against zero but not against a price‑only benchmark.
The paper proposes a centralized training with decentralized execution framework for multi‑agent UAV pathfinding in stochastic environments, leveraging multi‑agent reinforcement learning enhanced by model predictive control concepts. Agents communicate only with a centralized planner and use a distance‑weighted mean‑field observation to plan feasible paths, while multi‑step value convergence improves training efficiency. Experiments, including ablation and real‑robot studies, demonstrate the method’s effectiveness and generalization.
RealCLI is a new benchmark for evaluating human‑agent collaboration in command‑line interface (CLI) tasks. It unifies task execution and collaboration quality assessment through a five‑dimensional framework with 22 sub‑dimensions and 51 interaction templates. The benchmark offers two modes—REALCLI‑EXEC with full specifications and REALCLI‑COLLAB with a scripted user providing progressive feedback—and reveals significant capability gaps in current agents, especially in steerability, trust calibration, and safety management.
The paper evaluates confidence‑based routing of small‑to‑large language models across multiple tasks, languages, and model families, covering 72 configurations and over 42,000 predictions. It finds that confidence routing improves performance on reading comprehension and topic classification datasets, but offers little benefit on a knowledge‑reasoning benchmark where small models remain confidently wrong. The study also shows that global confidence routing can create large language‑specific performance gaps, which equal‑quota routing mitigates with minimal overall accuracy loss.
UniPolar is a new framework for restoring images and polarimetric data from color polarization filter array (CPFA) images that suffer from both device- and environment-induced degradations. It uses two complementary streams to reconstruct four polarization-angle images and predict Stokes difference components, coordinating them with a Degradation Conditioned Cross Domain Interaction module that exchanges frequency-selective evidence and applies condition-adaptive feature correction. The method also enforces physical consistency at the output level, aligning image-derived and directly predicted Stokes components while regularizing angular fidelity and the physical feasibility of degree of linear polarization (DoLP).
IMPLICIT-Bench is a benchmark of 1,831 prompt triplets that tests whether text‑to‑image models can override stereotypical defaults when given explicit requests. Each triplet contains a neutral prompt, a stereotype prompt, and an anti‑stereotype prompt, allowing measurement of a model’s default bias and its ability to follow contradictory instructions. Experiments on Qwen‑Image and Stable Diffusion 3 show that models often default to stereotypes, succeed less often on anti‑stereotype requests, and that prompt rewriting or embedding steering can modestly reduce bias when the relevant stereotype is known.
The paper introduces StratFL, a strategy for federated learning that counters covert model poisoning attacks by modeling the interaction between the server and attacker as an imperfect repeated game. It proposes new belief‑update algorithms for both parties and an adversarial search method to detect and mitigate intermittent poisoned updates. Experiments on MNIST and CIFAR‑10 show that StratFL improves accuracy by about 10% over existing defenses while demonstrating convergence of the belief updates across various security scenarios.
DeDiCa is a scalable framework for causal inference with discrete variables and hidden confounding. It uses a tensorized parameterization of causal Bayesian networks, allowing efficient maximum‑likelihood training on observational data while exactly marginalizing over hidden confounders. The model can estimate any (proxy‑)identifiable interventional query with CPU‑efficient training and millisecond‑scale inference, matching state‑of‑the‑art neural generative models in ATE estimation while requiring less memory and time.
The paper introduces a new spectral federated learning method that combines local updates and communication compression, addressing key bottlenecks in FL such as data heterogeneity and high communication overhead. It builds on the spectral optimizer Muon, which has shown superiority over Adam-like methods in centralized training, and adapts it to the federated setting. The authors provide rigorous convergence guarantees without requiring the restrictive bounded heterogeneity assumption and validate their approach on standard image classification benchmarks.
The paper investigates why large language models often answer factual questions incorrectly in closed‑book settings. It shows that many errors are due to first‑token recall failures: the model commits to an incorrect first token but can complete the correct answer once the right token is provided. Using a First‑Token Cueing probe, the authors demonstrate that correcting just the first token recovers a significant portion of errors across 16 open models and improves accuracy on multilingual benchmarks.
Neuton is a framework that automatically converts neuron‑level abstract interpreter specifications into tensor‑level executables, enabling the creation of new DNN certifiers without manual tensor programming. It introduces a stack‑based intermediate representation with shape analysis, domain‑specific rewrite rules for efficient tensor operations, and a novel sparse tensor format (gBCSR) tailored to certifier sparsity. The authors demonstrate that Neuton preserves neuron‑level semantics and evaluate it on eight certifiers across 32 DNN benchmarks, compiling four existing certifiers and producing executable implementations for four others, revealing new precision‑runtime trade‑offs such as SkipPoly matching DeepPoly’s precision while often running faster.
ODSS-Net introduces a framework for multi‑modal object re‑identification that separates shared structural information from modality‑specific cues. It employs a Shared Structural Feature Construction module to extract modality‑shared structure, an Orthogonal Specific Feature Learning module to isolate identity‑relevant local features, and a Sample Adaptive Mixture of Experts module to fuse these features based on shared structure. Experiments on three benchmarks confirm the effectiveness of this orthogonally decoupled representation learning approach.
Expand-and-Prune is a test‑time framework for candidate‑based end‑to‑end autonomous driving planners that dynamically expands and prunes the set of trajectory hypotheses. It adds new candidates guided by semantic driving intent and temporally propagated predictions, then uses uncertainty in intermediate trajectory scores to prune the expanded set. Across four planners and four benchmarks, the method consistently improves planning performance, achieving state‑of‑the‑art results on LangAuto and significant gains on Bench2Drive, Fail2Drive, and NAVSIMv2, especially in out‑of‑distribution scenarios.