GTPred is a new benchmark for geo‑temporal prediction that evaluates multi‑modal large language models (MLLMs) on 370 images taken across 120 years worldwide. It assesses predictions by matching both the year and a hierarchical location sequence, and includes annotated reasoning chains to test intermediate reasoning. Experiments on 15 MLLMs show that while visual perception is strong, models still lack world knowledge and geo‑temporal reasoning, and that adding temporal data improves location inference.
By Jinnao Li, Tingzhu Chen, Changbo Wang
The paper introduces Segmentation-Based Attention Entropy (SAE), a method that uses semantic segmentation to measure visual attention uncertainty in large vision‑language models (LVLMs). SAE provides a reliability score for detecting hallucinated objects and an attention‑adjustment technique that reduces hallucinations during inference. Experiments on public benchmarks and real quadruped robot scenarios demonstrate that SAE improves LVLM reliability without requiring additional training.
By Jiale Song, Jiaxin Luo, Xue-song Tang, Kuangrong Hao, Mingbo Zhao
RAVE (Re-Allocating Visual Attention) is a lightweight pair‑gating mechanism that adds a learned query‑key bias to pre‑softmax attention scores over visual keys, derived from pre‑RoPE query and key features. It requires no architectural changes to the backbone and can be trained end‑to‑end with the rest of the model. Across multiple multimodal benchmarks, RAVE improves standard attention by an average of 3 points, especially on perception‑intensive tasks such as multilingual OCR, chart understanding, document VQA, and scene text VQA.
By Xi Leng, Xinhong Ma, Ziqiang Dong, Feng Zhang, Xiaoying Tang, Yang Yang, Guanjun Jiang
The paper investigates whether fine‑tuning a language model erases previously embedded activation steering interventions that suppress refusals and encourage brevity. Across five instruction‑tuned models (3B–14B) subjected to non‑adversarial supervised fine‑tuning (SFT) and reinforcement learning from human feedback (RLHF), the authors find that the steering’s behavioural effect degrades when the fine‑tuning objective conflicts with the targeted behaviour, yet the underlying weight edits remain largely unchanged. Mechanistically, the steering vectors survive with minimal alteration, but functionally the steering is vulnerable and must be re‑validated after downstream training.
By Philipp E. Glass, Allan Tucker, Yongmin Li, Alina Miron
MyoMechanix is a multimodal dataset and framework for action quality assessment that incorporates muscle activity and other physiological signals alongside visual data. It contains over 7,500 samples of 20 weight‑loaded actions from 38 subjects, with synchronized RGB video, 3D pose, sEMG, and additional signals. The accompanying Fitness Knowledge Graph structures expert annotations into relationships among actions, phases, key steps, errors, and corrective feedback, enabling compositional scoring and interpretable assessment through the CUBIST engine. The project also introduces MyoMechanix‑AQA, MyoMechanix‑VideoQA, and a novel MyoMechanix‑Video2EMG task, demonstrating that multimodal sensing and structured representations improve performance, interpretability, and error attribution.
By Hao Yin, Paritosh Parmar, Lijun Gu, Lin Xu, Tianxiao Guo, Xiujin Liu, Tianyou Zheng, Yang Zhang, Weiwei Fu
The paper introduces PLUS, a framework that uses reinforcement learning to generate text-based summaries of individual users’ preferences, characteristics, and past conversations. These summaries condition a reward model, allowing it to predict personalized response preferences and improving reward accuracy by 11–77 % over the standard Bradley‑Terry model. PLUS demonstrates robust performance with new users and topics, achieves a 25 % improvement over existing personalized RLHF techniques, and enables zero‑shot personalization for state‑of‑the‑art models like GPT‑4.
By Hyunji Nam, Yanming Wan, Mickel Liu, Peter Ahnn, Jianxun Lian, Natasha Jaques
The paper investigates how different estimators of the reverse Kullback–Leibler (KL) divergence used as a regularization term in reinforcement learning (RL) training of large language models (LLMs) affect training stability and downstream performance. By analyzing gradient bias across various estimator configurations, the authors demonstrate that biased gradients can cause training instabilities, while unbiased configurations improve performance on both in‑domain and out‑of‑domain tasks. Experiments on Qwen2.5‑7B, Llama‑3.1‑8B‑Instruct, and Qwen3‑4B‑Instruct‑2507 confirm these findings and show that KL regularization also stabilizes off‑policy RL training in asynchronous setups.
By Vedant Shah, Johan Obando-Ceron, Vineet Jain, Brian Bartoldson, Bhavya Kailkhura, Sarthak Mittal, Glen Berseth, Pablo Samuel Castro, Yoshua Bengio, Esmeralda S. Whitammer, Moksh Jain, Siddarth Venkatraman, Aaron Courville
The paper introduces a diagnostic tool that measures how neural emulators of partial differential equations capture physical symmetries by evaluating the overlap of loss gradients along symmetry-related states. This metric probes the local geometry of the learned loss landscape and goes beyond traditional equivariance tests by directly assessing learning dynamics. Applied to autoregressive fluid flow emulators, the study shows that orbit-wise gradient coherence enables generalization over symmetry transformations and reveals when training selects a symmetry-compatible basin.
By James Amarel, Robyn Miller, Nicolas Hengartner, Benjamin Migliori, Emily Casleton, Alexei Skurikhin, Earl Lawrence, Gerd J. Kunde
The paper evaluates additive activation steering in chat and agent contexts, showing that the commonly used gain ratio (Δ_agent/Δ_chat) fails to reliably indicate potency and efficacy across multiple models and dose-response cells. By replacing the gain with a location metric, dEC50 (difference in EC50 between agent and chat), the authors demonstrate a more robust, two‑sided measure that consistently captures cross‑context shifts. The study also reports several refuted and unanswered claims, emphasizing that a single operating point cannot distinguish between displacement and gain effects.
By Lucas Pinto
The paper shows that AI‑text detectors, rather than learning a clear AI‑versus‑human boundary, amplify an inherited predictability axis that already exists in language models. This amplification causes detectors to over‑flag fluent, formal human writing while missing high‑temperature AI outputs, and the bias persists across languages, code, and detector architectures. A training‑free operator can relocate the bias but cannot erase it, underscoring that the unfairness is a structural cost of out‑of‑distribution generalization.
By Alexander Smirnov
IDEAlign introduces a new protocol for evaluating the similarity of large language model (LLM) annotations to expert judgments. It uses pick‑the‑odd‑one‑out tasks to capture expert similarity and benchmarks various similarity methods—including text embeddings, topic models, and LLM-as-a-judge—against these human ratings. Applied to educational datasets, the study finds that most metrics miss nuanced expert dimensions, with LLM-as-a-judge performing best yet still insufficient for full expert alignment.
By Hyunji Nam, Lucia Langlois, James Malamut, Mei Tan, Dorottya Demszky
TDFNet introduces a Tri-projection Deformable Fusion Network that uses equirectangular, cube map, and tangent projections to mitigate geometric distortions in panoramic salient object detection. It incorporates a cross-projection deformable attention module for geometry-aware sampling and a latitude-guided fusion module that balances ERP and CMP features using spherical latitude priors. The network’s three-branch encoding preserves global continuity, local detail, and boundary precision, improving detection performance over existing projection-based methods.
By Qiangqiang Zhou, Jiacong Yu, Jiawei Xu, Yong Chen, Xin Huang, Ping Li
THA-Flow is a conditional flow-matching model that generates 3‑D prosthesis geometry directly from preoperative CT scans for total hip arthroplasty. It uses separate AutoencoderKL models to compress bone anatomy and prosthesis shapes, and a 3‑D UNet to learn a flow from Gaussian noise to the prosthesis latent space conditioned on bone geometry. In a retrospective cohort of 1,355 hips, the model produced accurate acetabular and femoral geometries for 93.4% of cases, preserving component position and alignment while allowing limited local variation.
By Yiping Wang, Jie Li, Jingyu Shen, Liao Wang
The paper introduces a retrieval‑augmented multi‑agent framework that automatically generates instance‑specific evaluation rubrics for medical language models. By retrieving authoritative medical evidence, decomposing it into atomic facts, and combining these with user interaction constraints, the system produces fine‑grained criteria that outperform GPT‑4o on HealthBench and LLMEval‑Med. The generated rubrics also guide response refinement, improving medical LLM output quality by 9.2%.
By Yinzhu Chen, Abdine Maiga, Hossein A. Rahmani, Emine Yilmaz
The paper introduces DEDUCE, a three‑stage framework that turns large language models into proactive error correctors by detecting input fact errors, devising correction strategies, and delivering reliable answers. It also presents MisFactQA, a dataset of factual errors, and new metrics for robustness evaluation. Experiments on TruthfulQA, FalseQA, and MisFactQA show significant gains in accuracy and error correction across Qwen, LLaMA, and Gemma models.
By Ping Wang, Xiangguo Sun, Bingbing Xu, Guocong Li, Xiaofeng Meng
The study investigates how prior scores influence large language model (LLM) judgments in the LLM-as-a-Judge paradigm. By testing three prompt conditions—no metadata, revision framing, and anchored metadata containing prior scores—the authors find that prior scores systematically bias evaluations, shifting ratings toward those scores across 192,000 attempts. The bias also affects categorical decisions, blocking 48% of error corrections and flipping 10.18% of correct judgments, and is not mitigated by Chain-of-Thought or a warning, underscoring the need for careful context engineering.
By Ante Kapetanovic, Kemal Altwlkany, Andro Mercep, Tomislav Duricic, Emanuel Lacic
The paper investigates whether retrieval‑augmented generation (RAG) can uniformly correct factual errors in large language models (LLMs) by testing six LLMs on a benchmark of about 2,000 public companies. In a controlled factual QA setting, the authors evaluate the models under four conditions—no‑context, perfect context, misleading context, and distraction context—across four atomic attributes. Results show significant geographic disparities in baseline accuracy, and while perfect context improves performance, it does not eliminate these gaps; misleading context often leads models to copy incorrect information, and larger models only marginally reduce structural biases.
By Abhinav Havaldar, Enrico Santus
The paper investigates how refusal training shapes the internal geometry of language models, showing that activation updates from refusal-completion losses create a distinct low‑dimensional refusal subspace. In a case study on OLMo‑2‑0425‑1B‑Instruct, the authors link the brittleness of refusal directions to repetitive refusal prefixes and demonstrate that using diverse refusal starts can increase the stable rank of gradients, thereby hardening the model against vector‑ablation attacks. The work provides insights into the emergence of safety‑critical features and offers a potential strategy to strengthen refusal robustness.
By Andrey Labunets
The paper investigates hierarchical clustering under an individual fairness constraint that limits relative distortion within local k‑nearest neighborhoods. It formulates this as a feasibility problem over dominated ultrametrics, characterizes the minimal multiplicative slack needed, identifies a sharp local threshold, proves stability under bounded perturbations, establishes monotonicity in k, and demonstrates a Θ(log n) separation between local and global realizability. Experiments on synthetic and real‑world datasets corroborate the theoretical findings.
By Binita Maity, Shrutimoy Das
The paper introduces FairTPT, a fairness-aware test‑time prompt tuning method for vision‑language models like CLIP. It jointly minimizes target marginal entropy while maximizing spurious marginal entropy to reduce bias under subpopulation shifts. Experiments show that standard episodic test‑time adaptation can worsen disparities, but FairTPT outperforms existing debiasing methods while preserving overall performance.
By Yoann Launay, Parameswaran Kamalaruban, Tom Kempton, Stuart Burrell, David Sutton