Leaderboards, eval harnesses and ablations — the contested business of deciding which model is actually better.
Reinforcement learning (RL) is vital for optimizing video generation models, with a robust reward model (RM) serving as the cornerstone. However, existing video reward models often produce unstable sc...
Professional graphic design is a long-horizon agentic task in which structured, editable artifacts emerge from many interdependent actions, yet outcomes admit no reliable programmatic oracle. We intro...
Active reconstruction requires efficient active view selection to achieve high-quality reconstruction within limited onboard computational resources. Existing methods face challenges in adequately bal...
LaSR (Latent Speech Reasoning) is a new training paradigm for context‑aware speech recognition that uses a latent reasoning trajectory instead of explicit intermediate tokens. It aligns chain‑of‑thought supervision around the acoustic region of target words and introduces latent reasoning periods for grounding context and guiding transcription transitions. Experiments on Fun‑Audio‑Chat show that LaSR improves terminology recognition without added latency and outperforms standard fine‑tuning baselines, demonstrating the promise of latent reasoning for efficient, context‑aware speech assistants.
The paper introduces Decoupled Contrastive Decoding (DCD), a method that separates drafting from verification in contrastive decoding. DCD uses an expert‑aligned lightweight proposer for drafting and applies contrastive correction only during verification, preserving the vanilla contrastive decoding distribution. Experiments on 8B models show that DCD achieves 1.65–1.95× speedups in greedy decoding and reduces MMLU proposal‑path latency by 5–12× compared to amateur‑coupled approaches.
The paper introduces IMFD, an end‑to‑end multi‑face forgery detector that uses instruction‑based Large Vision‑Language Models (LVLMs). IMFD jointly localizes faces and predicts forgery labels in a single stage, explicitly incorporating predicted face bounding boxes into the textual instruction to improve grounding and detection. Experiments on converted multi‑face forgery datasets show that IMFD outperforms several state‑of‑the‑art methods.
STAR (Structure-aware Test-time Adaptation for diffusion-based light field Reconstruction) addresses the ill‑posed problem of reconstructing a light field from limited, noisy focal stack measurements by adapting a pre‑trained diffusion prior to each test scene. The method freezes the diffusion model and trains three lightweight adapters that jointly capture within‑view spatial details, cross‑view angular dependencies, and disparity across views. Experiments show that STAR surpasses state‑of‑the‑art approaches in both two‑ and three‑focal‑sheet settings while achieving faster inference times than methods that update parameters at test time.
SIMLIFE is a scalable platform that simulates long-term household life with rich visual observations, ground-truth action logs, and synthetic dialogues. It introduces the SimLife-BP benchmark, which tests long-context pattern understanding by requiring agents to infer latent behavioral rules from weeks or months of everyday observations across 106 episodes. The benchmark includes 1,439 question-answer pairs that probe direct, counterfactual, noisy, and inverse reasoning under varying rule hints.
The paper introduces a context‑aware neural grapheme‑to‑phoneme (G2P) system for unsegmented languages like Japanese, using a discriminative conditional random field over a word lattice built from dictionaries. It addresses data scarcity by generating over two million sentences with large language models. Experiments show the method surpasses traditional morphological analyzers and neural sequence models, achieving 99.62% target word reading accuracy and very low phoneme error rates on the Joyo‑Kanji‑Yomi benchmark.
The paper introduces F$^{2}$DR, a fine‑grained reward framework designed to evaluate end‑to‑end DeepSearch workflows, which involve planning, reflection, retrieval, and answer generation. F$^{2}$DR assesses workflows along three dimensions—Content, Trajectory, and Answer—to provide a comprehensive process‑level evaluation. The authors also present DeepSearch RM‑Bench, a benchmark that tests reward models in DeepSearch scenarios and shows strong discriminative power over existing open‑source models.
The paper introduces Bayesian Optimization (BO) techniques that incorporate rich auxiliary information—such as training curves, expert notes, images, and prior knowledge—using large language models (LLMs). Three new methods are proposed to integrate this auxiliary data into BO, and they are evaluated on hyperparameter optimization benchmarks and a real-world nuclear fusion task. The results show that these LLM-enhanced BO methods consistently outperform standard BO and existing LLM-based optimization approaches.
The paper "Benchmarking LLM Compliance with China AI Generated Content Regulations" presents a new framework for evaluating large language models (LLMs) against China’s AI‑generated content compliance rules. Using 2,303 questions across six dimensions—including 203 self‑constructed constitutional queries—the authors assess 20 prominent LLMs, finding that international models show high compliance, especially on ideologically aligned dimensions. The study establishes a regulatory benchmark that allows the global AI community to test both Chinese and non‑Chinese LLMs under unified, legally grounded compliance criteria.
The paper investigates whether language models’ self-reported confidence is meaningful without additional training. By evaluating three training‑free signals—direct verbalization, post‑hoc probability estimates, and agreement across multiple generations—on 100 TriviaQA questions, the authors find that direct verbalization alone achieves high AUROC scores (0.956 and 0.937) for correctness prediction, while agreement-based methods perform noticeably worse. Re‑eliciting confidence for the same answers shows modest score shifts and occasional decision flips, and an audit of biography claims reveals only a small confidence gap between supported and contradicted statements.
MERIT‑Rank is a reranking framework that integrates multiple reasoning trajectories to enhance text ranking robustness. It introduces a Multi‑Trajectory Reasoning Space (MTRS) to evaluate query‑document relevance from diverse perspectives and a joint reranker that consolidates these paths into a single ranking decision. The Progressive Rank Policy Optimization (PRPO) training scheme stabilizes reasoning trajectories and progressively improves ranking quality, yielding superior performance on both reasoning‑intensive and traditional retrieval benchmarks, with a 4B model outperforming larger 7B and 32B rerankers on BRIGHT.
The paper investigates a new failure mode of tool‑augmented large language model agents: calling non‑existent tools with arguments that do not match any declared schema. It introduces a five‑class taxonomy of tool hallucination, presents a training‑free closed‑world resolver that checks registry membership and signatures, and demonstrates that hallucinations persist across ten hosted models and various invocation surfaces, including the Model Context Protocol. The authors release a Hallucinated‑Tools Benchmark to enable comparison of resolver methods.
The paper introduces a dependency‑graph framework to formalize compositional reasoning in language models, defining three increasing levels of compositionality. Using data‑structure tasks with deterministic rewards, the authors observe a consistent asymmetry: training on decomposed skills does not reliably transfer to composed tasks, whereas training on composed tasks transfers more readily to decomposed ones. They provide a theoretical explanation for this asymmetry and evaluate its effects under length extrapolation, structural distribution shift, and transfer to unseen skills, concluding with a pilot study on real‑world tool‑calling benchmarks that suggests the phenomenon extends to practical settings.
The paper introduces LLM-as-an-Improver, a method that uses verification feedback to enhance the candidate set in verifier-based selection. It proposes Verify–Repair–Reselect (VRR), which keeps the initial winner, generates three complementary alternatives (repaired versions of the winner and runner‑up, and a new approach), filters invalid or duplicate candidates, and then reselects the final answer. Experiments on code‑generation and reasoning benchmarks show that VRR outperforms fixed‑pool selection and can recover correct solutions even when the initial pool is entirely wrong.
The paper introduces SIFT, a sample‑efficient framework for self‑improvement of coding agents that uses a fast tree‑search guided by an LLM‑as‑a‑judge signal. By performing pairwise comparisons of candidate patches and aggregating results with a regularized Bradley‑Terry model, SIFT limits expensive downstream evaluations to only the most promising nodes. The approach achieves higher coding performance on the Polyglot benchmark while reducing CPU hours, wall‑clock time, and API costs compared to prior tree‑search self‑evolution methods.
The paper introduces ScientistTwo, a fully autonomous multi‑agent framework that takes a scientific problem, establishes baselines, generates hypotheses, and coordinates specialized agents to conduct an end‑to‑end discovery cycle without human intervention. It rigorously tests and refines its methods through automated experiments, ablation studies, and a closed‑loop peer‑review engine. Benchmarking against top conferences (ICLR, ICML, NeurIPS) shows that ScientistTwo produces expert‑level, publishable papers and codebases that outperform human state‑of‑the‑art models and receive higher review ratings under automated AI review.
TorchCraft is a unified binder‑design framework that optimizes sequence logits using a frozen all‑atom structure predictor. It integrates confidence, contact, geometric, and sequence‑prior objectives within TorchFold to design minibinders, framework‑conditioned VHHs, cyclic peptides, and ligand‑binding proteins. Using pretrained AlphaFold 3 weights, TorchCraft produced experimentally validated binders across four targets without post‑hoc redesign, and computational tests confirmed its applicability to cyclic peptides and ligand‑conditioned pocket design.