The paper introduces Mentalized Adversarial Self-Play (MASP), a framework for training proactive dialogue agents by modeling the user as a rationally resisting participant with an evolving mental state. MASP separates teacher and student mentalization, alternates self‑play updates for system and user policies, and treats robustness to rational resistance as a training objective. Experiments on proactive dialogue benchmarks show MASP improves success rate, rationality‑constrained success, reduces average turns, and performs better against hard users compared to SFT and Planner‑only approaches.
By Zanwu Liu, Dan Zhang, Xiaodi Wang, Yongbo Mao, Xiaoyang Guo, Yonglin Wu, Peiyue Jiang, Yangyang Wei, Zecheng Lin
CLUE is a two‑stage training pipeline that teaches models to discover and use social context cues within a conversation for pragmatic understanding. It combines supervised evidence reasoning with a GRPO reward that credits new, evidence‑supported social inferences while penalizing excessive reasoning effort. Experiments show that CLUE improves weighted F1 scores by 17.3% in‑domain and 9.3% out‑of‑domain, reduces reasoning tokens by over 80%, and aligns its explanations more closely with human‑selected conversational evidence.
By Chengfei Wu, Zhaoqing Wu, Dan Goldwasser
EVA-Cap is a new framework for audiovisual video captioning that breaks captions into atomic audiovisual events, turning unstructured text supervision into event‑centric alignment. The model is trained with supervised fine‑tuning and Group Relative Policy Optimization on a self‑curated dataset, EVA‑Data, which applies event‑centric quality control. Experiments show EVA‑Cap outperforms prior methods on event‑centric metrics (e.g., +9.0% on ChronusAV) while remaining competitive on general captioning benchmarks (e.g., +3.3% on WorldSense).
By Jiapeng Shi, Qiuxia Hou, Jiaqi Leng, Junke Wang, Yanhao Zhang, Haonan Lu, Ziyi Ye, Zuxuan Wu
The paper shows that large language models can infer ethnicity from names with higher accuracy than Bayesian Improved Surname Geocoding (BISG), achieving up to 84.7% accuracy on balanced samples from Florida and North Carolina voter files. Six models, including Gemini 3 Flash, GPT‑4o, and open‑source alternatives, were tested, with extended reasoning and metadata such as party registration further improving performance. The method also reduces income bias present in BISG and is validated across diverse populations in India, Uganda, Nepal, Armenia, Chile, and Costa Rica, with fine‑tuned small transformer models surpassing BISG while allowing local deployment at no cost.
By Noah Dasanaike
The paper introduces Step-level Self-Consistency Group Relative Policy Optimization (SSC‑GRPO), a method that assigns rewards to individual steps in a reasoning trace by evaluating self‑consistency across multiple rollouts. It targets Context‑Sensitive Factual Hallucinations—errors that arise when a model has the correct knowledge but produces false statements due to contextual interference. SSC‑GRPO outperforms previous approaches on mathematical reasoning benchmarks and hallucination leaderboards, offering a new way to detect and reduce hallucinations in large language model reasoning.
By Xiaomeng Hu, Jiaqi Hu, Hao Chen, Qi Zhang, Zhanming Shen, Wentao Ye, Junbo Zhao
The paper examines how large language model (LLM)-generated texts affect term-based retrieval models such as BM25. It finds that LLM outputs have smoother high-frequency and steeper low-frequency Zipf slopes, higher term specificity, and greater document-level diversity, traits that align with LLMs’ focus on reader experience. The study shows that term-based models prioritize documents whose term distributions match the queries, rather than showing an inherent bias toward the source of the text.
By Wei Huang, Keping Bi, Yinqiong Cai, Wei Chen, Jiafeng Guo, Xueqi Cheng
FocusOPD introduces a structure‑guided on‑policy distillation method that adds token‑wise directional feature alignment to deep Transformer layers during early denoising steps. This approach addresses the mismatch between teacher and student representations that persists after standard OPD distillation. Experiments on rectified‑flow backbones show that FocusOPD improves performance on GenEval, PickScore, and LAION‑Aesthetic benchmarks, and ablations confirm that selective deep and early supervision is more effective than full‑network alignment.
By Xinrui Lin, Shumin Wang, Suqi Song, Fan Li, Yanyong Zhang
The paper presents a method for synthesizing contrast‑enhanced MRI (CE‑MRI) from non‑contrast sequences using a novel Mixture of Feature Difference (MoFD) module. MoFD dynamically weights and aggregates features from individual sequences and their explicit differences, improving image fidelity and perceptual quality over baseline approaches. Experiments on UCSF‑PDGM and UPenn‑GBM datasets show that the synthesized images reconstruct tumor regions well and yield segmentation performance comparable to real CE‑MRI, with regression‑based training achieving higher structural fidelity.
By Jonghun Kim, Junmo Kwon, Hyunjin Park
PhyMix introduces a Physics Evaluator that checks geometric priors, contact, stability, and deployability through nine sub-constraints, establishing a benchmark for physical consistency in single-image 3D indoor scene generation. The framework uses implicit alignment via Scene-GRPO and explicit refinement through verifier-guided agentic repair to incorporate evaluator feedback during training and inference, improving physical plausibility while preserving scene content. Experiments demonstrate enhanced reconstruction quality and physical plausibility, with qualitative results on stylized and real-world images showing the method’s broader applicability.
By Dongli Wu, Jingyu Hu, Ka-Hei Hui, Xiaobao Wei, Chengwen Luo, Jianqiang Li, Zhengzhe Liu
The paper introduces BiShield-TEE, a bilateral obfuscation framework for on-device large language models that jointly transforms input and output dimensions to create ambiguous solutions, thereby protecting the model from reverse engineering. It identifies a vulnerability in existing unilateral obfuscation methods—called the Solution Determinacy Property (SDP)—which allows attackers to recover the model’s true key and achieve high attack accuracy. Experiments demonstrate that BiShield-TEE limits attack success to the baseline black‑box level while improving inference efficiency on real hardware.
By Zhiyuan Wang, Yuezhi Zhou, Ziqi Zhang, Zhu Zhang, Zhixing Tan, Yongheng Deng
The paper introduces Stochastic Truncated On-Policy Distillation (STOPD), a method that speeds up on-policy distillation by randomly truncating student rollouts and correcting sampling bias with blockwise inverse-probability weighting. STOPD retains teacher feedback for every generated token and preserves the expected gradient of vanilla OPD while providing theoretical error bounds for common optimizers. Experiments show that STOPD cuts generated tokens by 25–48% and achieves 1.3–2.2× faster training without sacrificing accuracy.
By Zhengxin Zhang, Maosen Tang, Ning Wang, Changhao Li, Jingren Zhou, Bo Dai, Sainyam Galhotra, Alex Townsend, Claire Cardie
The paper introduces Router Drift Control (RDC), a method for mitigating off‑policy degradation in Mixture‑of‑Experts reinforcement learning. RDC uses the disagreement between rollout and training routers to detect stale routing decisions and selectively rejects high‑drift responses, thereby avoiding the gradient bias introduced by Rollout Routing Replay (R3). Experiments show that RDC improves validation accuracy, especially in asynchronous training settings, outperforming both R3 and Truncated Importance Sampling (TIS).
By Liangyao Jin, Zhixin Wang, Liming Liu, Mingjun Zhang, Zhengzhou Zhu, Fangcheng Fu, Jinlong Hou, Yuan Cheng, Xiaohe Hu, Ping Zhang, Siyuan Feng
GeoState introduces a view‑decoupled geometry representation for predictive 3D world modeling, separating scene‑state evolution from camera‑dependent observation. It aggregates multi‑view geometry foundation model features into a scene tokenizer, then uses a geometry decoder with camera rays to query the state. An autoregressive world model predicts future scene states while modeling camera motion separately, enabling consistent decoding under any camera and maintaining a single context‑derived scale throughout the rollout.
The paper introduces a dual‑view approach for time‑series language models, combining line‑plot visualizations that preserve absolute measurements with normalized delay embeddings that remove amplitude scaling and offsets while exposing temporal structure. By fusing these complementary representations through visual tokens and multi‑level feature fusion, the model achieves significant performance gains on several benchmarks, improving TSRBench accuracy from 44.5% to 53.2% and outperforming dedicated TSLM baselines. The results demonstrate that leveraging both absolute and relative temporal cues enables more effective reasoning over time‑series data.
AVC-Reward introduces an event‑centric framework for evaluating audio‑video consistency in joint generation models. It separates audio and visual events, assesses them independently, cross‑validates cross‑modal relations, and aggregates results into a three‑level metric covering global narrative harmony, event‑level synchronization, and fine‑grained attribute alignment. The resulting AVC‑Reward model aligns closely with human judgments, offers interpretable failure localization, and enhances consistency when used as a reinforcement‑learning signal.
Multimodal large language models (MLLMs) can act unsafely when individually benign images and texts are interpreted jointly. Beyond implicit reasoning risk, where reasoning drifts toward harm, we iden...
Names carry demographic indicators such as age, gender, and race, which makes them a subtle but strong target of bias in large language models (LLMs). In this work, we present an evaluation of name-ba...
Language-guided 3D affordance grounding aims to localize functional regions on 3D objects according to interaction intents. However, existing methods exhibit poor generalization to unseen objects. One...
Despite the remarkable generalization capabilities of Large Language Models, generating clinically accurate and structured radiology reports from medical images remains a significant challenge. Curren...
Language embeddings are increasingly used in recommenders to represent item content, but their privacy risks remain poorly understood. We study whether item embeddings can reveal item identities when...