The paper investigates demographic bias in large language model (LLM) predictions of ICU mortality and introduces Case-Based Prompting (CAP), a training‑free method that adds clinically similar historical cases as contextual evidence. CAP improves predictive performance on the MIMIC‑IV dataset, raising AUROC from 0.806 to 0.873 and AUPRC from 0.497 to 0.694, while reducing subgroup disparities for sex and White‑Black comparisons. The study also shows that simply removing demographic information does not uniformly enhance fairness or accuracy, and that feature‑dependence patterns remain consistent across groups.
By Gangxiong Zhang, Yongchao Long, Yuxi Zhou, Yong Zhang, Shenda Hong
MosaiChunk introduces a spatio-temporal memory mechanism that composes a mosaic of selected historical key-value entries across space and time to improve long-horizon autoregressive video generation. By keeping the video generator frozen and learning a lightweight router to select historical sections within a fixed memory budget, the method enables the generator to recover fine-grained visual details when objects or scenes reappear. Experiments on the RememBench benchmark demonstrate that MosaiChunk consistently outperforms sliding-window inference and whole-chunk retrieval in both text-to-video and image-to-video settings.
By Yiwen Zhang, Haocheng Xi, Michael Tian-Yue Liu, Alexei A. Efros, Hadar Averbuch-Elor, Qianqian Wang, Haiwen Feng
LangMap introduces a human‑verified benchmark for language‑conditioned goal navigation (LGN) that spans four hierarchical semantic levels—scene, room, region, and instance—within real‑world indoor 3D scans. The dataset, built on HM3D, contains 18K tasks with concise and detailed descriptions for 414 object categories, and its contrastive annotation protocol ensures high‑quality, discriminative region and instance labels. Evaluation shows that LangMap’s descriptions improve text‑to‑view matching accuracy by 23 points over GOAT‑Bench and achieve a 92.5% unique‑and‑correct match rate in an independent human audit, while a proposed RGB‑only baseline, PlaNaVid, attains top‑tier success rates without depth or 3D scene representations.
By Bo Miao, Weijia Liu, Jun Luo, Lachlan Shinnick, Jian Liu, Thomas Hamilton-Smith, Yuhe Yang, Zijie Wu, Vanja Videnovic, Feras Dayoub, Anton van den Hengel
The paper introduces a 3D keypoint detector that replaces traditional heuristic post‑processing with a learned suppression module. This module, implemented as a graph neural network over candidate keypoints, learns to keep, suppress, or relocate points, and is built on a Point Transformer backbone extended with a directional graph neural network. Experiments show the approach outperforms DBSCAN and greedy non‑maximum suppression, surpasses KeypointDETR on most KeypointNet categories, and achieves top performance on the Building3D benchmark while remaining competitive on the Tallinn split.
By Batuhan Arda Bekar, Can Sar{\i}, H\"useyin Can G\"ulkan, Bar{\i}\c{s} \"Ozcan
The paper introduces REHAB26-ViewAngles, a dataset of correct and incorrect rehabilitation exercise videos captured from many camera angles, and a new separability metric to evaluate pose‑estimation algorithms. It evaluates single‑camera 2D and 3D pose estimation and four multi‑camera approaches, showing that an optimally placed 2D camera can boost separability by 16.9% over a frontal view and that combining two views can improve accuracy by up to 13.1%. These findings provide concrete guidance for setting up camera systems in home and clinical rehabilitation monitoring.
By Miriama J\'ano\v{s}ov\'a, Andreas Lang, Petra Budikova, Jan Sedmidubsky
The paper introduces AEGIS, a buffer‑free, layer‑wise orthogonal gradient projection framework that enables continuous fine‑tuning of Vision‑Language Models for robotic manipulation while protecting pre‑trained visual reasoning. AEGIS uses pre‑training activation statistics as static anchors and applies a Wasserstein‑2 transport penalty to generate an anchor‑restoration gradient, followed by a dual‑backward pass that orthogonalizes task gradients against this restoration vector. Experiments on the LIBERO manipulation benchmark with PaliGemma2‑3B show that AEGIS preserves Visual Question Answering performance and baseline loss while achieving continuous action convergence, all without replay buffers, teacher models, or co‑training data.
By Guransh Singh
The paper introduces memetic trojans, a new class of attacks that embed adversarial payloads within socially contagious content shared by large language model agents. Unlike traditional agent worms that rely on malicious retransmission, memetic trojans exploit agents’ natural tendency to share content that aligns with their internal motivations, leading to high virality and significant amplification of exposure. Experiments on the Moltbook platform show that such contagions can be retransmitted in about 50% of subsequent posts and upvoted 2.5 times more than average, with simulations indicating up to a 3.19x increase in expected exposure across the network.
whyItMatters":"The study reveals that endogenous social transmission is a distinct security vulnerability in multi‑agent systems, implying that defenses must address network-level dynamics rather than just prompt‑injection detection or agent compromise prevention."
By Birk Torpmann-Hagen, Finn Schwall, Leon Moonen
The paper "Where's Waldo? Query-language Preference under Cross-lingual Knowledge Disparities" introduces the Waldo benchmark, a multilingual QA dataset built from Wikipedia that focuses on knowledge gaps and conflicts across languages. It evaluates eight models in five languages and finds that when a fact is missing in one language, models tend to use evidence from the other language, but when conflicting accounts exist, responses align strongly with the query language, leading to different answers for semantically identical questions. The study also explores mitigation strategies, including ablating attention heads and LoRA-based training, which can reduce the preference gap by up to 61.5%.
By Dayeon Ki, Ruochen Zhang, Silviu Cucerzan, Ryen W. White, Ning Gao
The paper presents a matched‑budget audit framework for evaluating recaptioned image‑text supervision distributions, which are created by applying a captioning policy, captioner, and source corpus to generate dense descriptions for text‑to‑image training. The framework operates under a fixed text budget of 64 tokens and produces a five‑axis profile covering prompt coverage, image‑conditioned faithfulness, and caption surface quality, using controllable basic units as a common claim metric. The authors apply the framework to seven paired comparisons across five public corpora, demonstrating improvements in supported claim units and revealing consistent long‑vs‑dense trade‑offs across different judges and budgets, and they release the audited multi‑source recap corpus and audit artifacts.
By Giyeong Oh, Junghun Park, Yuhan Bae, Youngjae Yu
The paper investigates how the choice of reconstruction loss, specifically mean squared error (MSE), affects optimization in sequential post‑training quantization (PTQ) of large language models (LLMs). It identifies an Optimization Imbalance where MSE causes reconstruction loss magnitudes—and thus gradient magnitudes—to vary dramatically across quantization stages, leading to uneven parameter updates. The authors propose that root mean squared error (RMSE) variants, which implicitly normalize gradients, can decouple optimization strength from loss scale and outperform MSE as a drop‑in replacement.
By Chao Li, Shigeng Wang, Anbang Yao
The paper investigates whether reviewers with limited expertise can reliably determine if a coding agent’s patch resolves a bug, using 411 execution-labeled traces from three agents and 101 controlled cases. It finds that providing structured but unchecked evidence can improve defect detection and reduce over-rejection, and that a frozen cascade of static errors and generated tests achieves high coverage and catch rates on held‑out GPT‑5.4 and Gemini traces. Reviewer size alone does not predict quality, and the study highlights the challenge of creating reliable checks without official tests.
By Junyu Guo, Shangding Gu, Ming Jin, Javad Lavaei
The paper "It Takes Workflows to Evolve Better Workflows" introduces FloWright, a method that uses a hierarchical, structure‑aware reward system to allow one or more roles in a multi‑agent workflow to self‑evolve without extra models or data. It also proposes DataWright, an adaptive data hardening technique that transforms existing datasets into more challenging workflow‑level tasks. Experiments on document, slide, chart, code, math, and finance tasks show that small open models trained with FloWright can improve performance by up to +7.41%, with co‑evolving multiple roles yielding the largest gains.
"whyItMatters":"The work demonstrates that optimizing beyond the workflow generator—by enabling multiple agents to co‑evolve—can substantially enhance the effectiveness of multi‑agent workflows for complex real‑world tasks."
By Xuehang Guo, Haoyu Wang, Haifeng Chen, Yangyi Chen, Zhenhailong Wang, Qingyun Wang
OrbitTAMP introduces a hierarchical framework that grounds large language model (LLM) reasoning in spacecraft task-and-motion planning (TAMP) for rendezvous and proximity operations. The system maps natural‑language commands to partial mission specifications, resolves unspecified decisions with domain‑specific planners, and then optimizes trajectories for dynamic feasibility. Experiments show significant improvements in intent recovery, achieving 98% exact recovery with frontier LLMs and enhancing performance further with verifier‑guided revision and behavior‑plan search.
By Yuji Takubo, Daniele Gammelli, Marco Pavone, Simone D'Amico
The paper presents an automated framework for building ontology networks by discovering typed semantic links between concepts across domains. It uses domain‑adapted DistilBERT embeddings, clustering to prune candidates, and GPT‑4o for generating interpretable relationships, reducing 800k raw pairs to 95k high‑quality candidates. Human validation on 429 links shows an 80.19% precision and 0.890 F1, outperforming five similarity‑based baselines and demonstrating the need for LLM reasoning in ontology link construction.
By Nouha Hayouni, Sheeba Samuel, Alsayed Algergawy
arXiv:2610. 01493v1 Announce Type: cross Abstract: Iterative fine-tuning on synthetic data causes \emph{model collapse}: output diversity narrows as rare patterns are progressively lost, a signature most visible as phrase-level repetition.
By Lewis Mitchell
arXiv:2610. 01871v1 Announce Type: cross Abstract: Multimodal Retrieval-Augmented Generation (MRAG) has emerged as a reliable and cost-effective technique of grounding the generative capabilities of Multimodal Large Language Models (MLLMs) into relevant, up-to-date, external knowledge.
By Maria Carmen Jica, Ali Satvaty, Suzan Verberne, Fatih Turkmen
MoLE introduces a Mixture of Latent Experts framework that encourages distinct latent tokens to extract complementary visual information, rather than sharing visual evidence. By isolating latent visual experts during evidence extraction and aggregating their outputs with dedicated summary experts, MoLE achieves higher performance on visual reasoning benchmarks compared to existing latent reasoning methods. The approach uses a two-stage training pipeline that first forces visual evidence through the latent pathway and then restores direct visual access, without requiring predefined expert roles or intermediate visual targets.
By Yingcheng Liu, Tianyi Jiang, Yujuan Ding, jiangbo Ai, Xun Jiang, Guoqing Wang, Wei Ye, Yi Bin
Reconstruct, Practice, Go Real (RPG) is a framework that enables autonomous improvement of robot execution systems without updating model weights. RPG identifies manipulation capabilities from offline data, creates simulation practice tasks, and uses feedback to diagnose failures, develop new symbolic skills, refine existing ones, and revise the system prompt. Across 22 manipulation tasks, RPG raises task success from 28.6% after the first practice round to 95.0% after 15 rounds, outperforming baselines and achieving perfect success on 30 physical trials after calibration.
By Yen-Jen Wang, Haozhe Jiang, Shuying Deng, Haoru Xue, Weirui Ye, Rocky Duan, Nika Haghtalab, S. Shankar Sastry, Pieter Abbeel, Haozhi Qi
The paper investigates automated reward systems for training language models in conversational humor, examining how reward exploits can undermine intended behavior. It evaluates two reward approaches—an embedding-based surprise reward and an audience-model laughter prediction—showing that each can be tricked by word shuffling or laughter cues, respectively. Countermeasures such as fluency filtering and cue normalization mitigate some attacks but also risk rejecting genuine witty responses, highlighting the difficulty of designing robust rewards.
By Sam Larson
The study investigates how large language models (LLMs) encode persona-dependent preferences by training linear probes on residual-stream activations of Gemma‑3‑27B and Qwen‑3.5‑122B. It identifies a genuine preference vector that tracks the model’s task choices across various prompts and shows that steering along this vector can causally control pairwise choices. The research also finds that some preference information transfers between different personas, including a case where an evil persona’s preferences anti‑correlate with those of a helpful assistant.
By Oscar Gilg, Pierre Beckmann, Daniel Paleka, Patrick Butlin