The paper introduces a reasoning‑aware compression framework for Large Reasoning Models (LRMs) that identifies vulnerable reasoning circuits and selectively restores them to FP16 to avoid performance loss. By benchmarking INT4 quantization across five reasoning tasks (GSM8K, FOLIO, MATH‑500, ProofWriter, MuSiQue) and measuring GPU energy, the authors find that uniform quantization can actually increase energy consumption and that vulnerability varies by task and architecture. Selective compression yields Pareto‑optimal trade‑offs, achieving significant energy savings while improving or matching accuracy on held‑out data.
By Leonard Twagirayezu, Prasenjit Mitra
The paper introduces Dynamic DAE Guardrails (DSG), a method that uses Dynamic Sparse Autoencoders to perform precision unlearning in large language models. DSG leverages principled feature selection and a dynamic classifier to target activation-based unlearning, outperforming existing gradient‑based methods in terms of computational efficiency, stability, sequential unlearning, resistance to relearning attacks, data efficiency, and interpretability.
By Aashiq Muhamed, Jacopo Bonato, Mona Diab, Virginia Smith
The paper introduces Contrastive Projection, a method that reads a transformer’s internal states by differencing the hidden states of two closely matched prompts and projecting the difference through the unembedding layer. This approach cancels shared components and highlights the distinctions between prompts, effectively revealing steering vectors and domain-to-domain mappings such as metaphor. The technique is training‑free, operates at every position, sub‑layer, and head, and has been validated across multiple architectures and initialization seeds.
By Olli Tuomi
arXiv:2609.09206v1 Announce Type: cross
Abstract: Multimodal Large Language Models (MLLMs) often struggle with hallucinations, thus hindering their reliable practical applications. Existing attention...
By Meng'en Qin, Junye Chen, Jucheng Liu, Youlu Xing, Song Wang, Ruize Han
The paper demonstrates that fine‑tuning large language models on a single author’s works can trigger the models to reproduce large verbatim excerpts from copyrighted books, even when prompted only with semantic descriptions. Experiments on GPT‑4o, Gemini‑2.5‑Pro, and DeepSeek‑V3.1 show up to 85‑90% recall of held‑out books, with spans exceeding 460 words, and this effect generalizes across authors and model providers. The findings suggest that fine‑tuning reactivates latent memorization from pre‑training, revealing a widespread vulnerability in industry models.
By Xinyue Liu, Niloofar Mireshghallah, Jane C. Ginsburg, Tuhin Chakrabarty
PopResume is a population‑representative resume dataset designed for causal fairness auditing of large language model (LLM) and vision‑language model (VLM) resume screeners. It grounds fairness evaluation in real population statistics and preserves natural attribute relationships, enabling path‑specific effect (PSE) analysis that separates business‑necessity from redlining pathways. Using PopResume, the authors evaluated eight models on 60.8K resumes across five occupations and uncovered five discrimination patterns that aggregate metrics missed, demonstrating the value of causally‑grounded auditing.
By Sumin Yu, Juhyeon Park, Taesup Moon
The paper investigates how different components of a graph retrieval‑augmented generation pipeline affect large language model performance on knowledge‑graph question answering. It examines four variables—whether the answer path is included, the syntax of triples, the order of triples, and the subgraph size—across six LLMs and two benchmarks. The study finds that including the answer path is crucial, while the grounding instruction dramatically reduces accuracy when no facts are provided, and that syntax, order, and subgraph size have negligible measurable impact at multi‑hop depth.
By Arquimedes Canedo
EvidenceNet is a disease‑specific dataset that transforms full‑text biomedical literature into structured evidence records and graph representations, preserving study design, provenance, and quantitative support. Using an LLM‑assisted pipeline, it extracts experimentally grounded findings, normalizes entities, scores evidence quality, and links related records via typed semantic relations. The released subsets—EvidenceNet‑HCC and EvidenceNet‑CRC—contain thousands of evidence records and richly connected graphs, with high extraction and relation‑type accuracy, enabling retrieval‑augmented question answering and graph‑based tasks such as link prediction and target prioritization.
By Chang Zong, Jinyu Chen, Sicheng Lv, Si-tu Xue, Huilin Zheng, Jian Wan, Lei Zhang
The paper introduces an attack that reconstructs text generated by locally hosted large language models by monitoring CPU cache activity during detokenization. It uses Flush+Reload on shared tokenizer code to time decoding, then Prime+Probe to capture token‑dependent cache traces, followed by a clustering‑and‑language‑model pipeline to recover the output text. The method is evaluated across various datasets, hardware, inference frameworks, and model families, successfully retrieving semantically accurate outputs from real‑world local LLM deployments, including agentic systems.
By Roy Weiss, Benyamin Konstantinov, Eitam Sheetrit, Tomer Simon, Yisroel Mirsky
This paper introduces a deletion interface for a pretrained language model, measuring how effectively deleted records are removed from the model’s memory. By retrofitting a support‑vector memory gate into the global attention layers of a frozen Gemma 3, the authors show that deletions can be performed without altering weights and that the resulting state is close to a reference state that never stored the record. Experiments on 4B‑parameter models demonstrate low perplexity impact and strong evidence that deleted content is hard to recover, while larger or smaller models fail to achieve the same guarantees.
By Vishwajith Ramesh
The study investigates how a single politically charged framing term can influence large language models (LLMs) during translation tasks. By prompting eight models from Western, Chinese, and European origins to translate culturally attributed recipes across 17 languages under four framing conditions, the authors find that models resolve ambiguity rather than decline, with distinct behavior patterns tied to model families. Sensitivity to framing terms is consistent, showing that even subtle variations can modulate LLM behavior, raising concerns about implicit political judgments in translation contexts.
By Svetlana Gorovaia, Angelica Henestrosa, Ivan P. Yamshchikov
The paper introduces AMA, a framework that uses multiple agents—Constructor, Retriever, Judge, and Refresher—to manage memory for large language model agents. AMA’s hierarchical memory design dynamically adjusts retrieval granularity to match task complexity, while the Judge and Refresher ensure relevance, consistency, and timely updates. Experiments on long-context benchmarks show AMA outperforms existing baselines and cuts token usage by about 80% compared to full-context approaches.
By Weiquan Huang, Zixuan Wang, Hehai Lin, Sudong Wang, Bo Xu, Qian Li, Beier Zhu, Linyi Yang, Chengwei Qin
TamilEOT is a new dataset and two audio‑only models for detecting semantic end‑of‑turn in Tamil telephone speech, comprising 18,485 labeled turn boundaries from 116 real conversations. The models, fine‑tuned from Smart Turn v3, achieve 83.71% and 86.13% accuracy on a held‑out set, with ROC‑AUC rising from 0.751 to 0.921, and run in under 150 ms on a laptop CPU. The paper also reports the cost of building the dataset, the accuracy of rule‑derived labels versus human agreement, and the impact of encoder capacity on performance.
whyItMatters":"The study provides the first publicly available Tamil end‑of‑turn detection resource, enabling more accurate voice agents for South Indian languages and demonstrating efficient, low‑latency models that outperform zero‑shot baselines."
By Santhoshkumar V
SpatialBlock introduces a synthetic dataset of 15,000 block‑stacking problems designed to improve spatial intelligence in Large Vision‑Language Models (LVLMs). The dataset covers 3D‑to‑2D projection, viewpoint transformation, and structural combination, and uses controlled color modulation to encourage anchor‑based reasoning. Experiments show that LVLMs trained on SpatialBlock outperform baselines and generalize to real‑world spatial tasks, despite the dataset’s synthetic and compact nature.
By Soohyun Ryu, Sohee Kim, Eunho Yang
The paper introduces AV-STE, a modular streaming audio‑visual front‑end that enhances corrupted semantic speech tokens using noisy audio and lip video before they reach a frozen speech LLM. By preserving the downstream dialogue model’s pretrained conversational abilities, AV‑STE improves response coherence from 1.42 to 1.91 in same‑dataset speaker interference scenarios while maintaining turn‑taking behavior. These gains also transfer to out‑of‑domain Seamless Interaction.
By Bella Godiva, Yeonju Kim, Yong Man Ro
ReST‑RL introduces a unified Reinforced Self‑Training (ReST) policy‑value framework that enhances large language model (LLM) reasoning by combining an optimized ReST‑style GRPO algorithm with a value‑guided search (VM‑MCTS). The ReST‑GRPO component reshapes trajectory distributions to increase reward variance and expose policies to more informative partial states, improving training efficiency. VM‑MCTS trains a Value Model from self‑collected Monte‑Carlo Tree Search targets and uses it during inference to provide precise process signals and verification scores, boosting reasoning accuracy across coding benchmarks and out‑of‑domain math and science tasks.
By Sining Zhoubian, Dan Zhang, Jie Tang
Osprey is a target‑agnostic pre‑training method that bootstraps draft models for speculative decoding from existing small language models. By pruning to a shallow backbone, restoring language‑modeling capability with next‑token pretraining, and adapting via vocabulary alignment and distillation, Osprey reduces per‑target work to a lightweight adaptation step. Experiments show that a single Osprey backbone improves mean acceptance length by up to 22.7% and increases tokens per second by 17.5% across several large target models, especially on out‑of‑domain and multilingual data.
By Fengxiang Bie, Yuqing Jian, Yifan Yu, Zhongzhu Zhou, Zelei Shao, Ben Athiwaratkun, Shuaiwen Leon Song, Chenfeng Xu, Xiaoxia Wu, Tianyi Zhang
The paper introduces REASONS, a benchmark of 12,723 sentence-level citation instances across 12 arXiv subject categories, to evaluate scientific citation attribution under different evidence conditions. It proposes a dual-metric framework—Abstention Rate (AR) and Hallucination Rate (HR)—to balance reliability and responsiveness. Experiments with proprietary and open-source LLMs across various prompting and retrieval settings show that advanced Retrieval-Augmented Generation (RAG) reduces hallucinations but increases abstention, while adversarial metadata can push hallucination rates above 85%. Human evaluation confirms a high ratio of factual hallucinations to acceptable paraphrases, underscoring the need for systems that can appropriately abstain under uncertainty.
By Deepa Tilwani, Yash Saxena, Seyedali Mohammadi, Ankur Padia, Edward Raff, Amit Sheth, Srinivasan Parthasarathy, Manas Gaur
The study investigates how users interact with ChatGPT for code generation beyond simple function-level tasks, focusing on project-level benchmarks that involve multi-class dependencies. A user study with 36 participants examined prompting patterns, screen recordings, and chat logs to identify Human‑LLM Interaction (HLI) features that influence productivity. The results highlight three consistently supportive HLI features, five guidelines to boost productivity, and a taxonomy of 29 runtime and logic errors with mitigation strategies.
By Sangwon Hyun, Hyunjun Kim, Jinhyuk Jang, Hyojin Choi, M. Ali Babar
The paper argues that large language models cannot achieve perfect reliability for any task, even with unlimited scale. It establishes that each generative task has an inherent reliability ceiling set by how much output uncertainty can be resolved from observable context, with a resolvable part that can be improved by more context and a subjective part tied to task ambiguity. The authors derive a scaling law showing that performance is limited by the scarcer resource—either training data or model capacity—and explain how this law explains phenomena such as retrieval augmentation and catastrophic forgetting.
By Subhabrata Majumdar