The paper introduces STAG, a post‑hoc framework that provides token‑level spectro‑temporal grounding for captions produced by audio‑based multimodal large language models (MLLMs). STAG estimates temporal support for each token via vocabulary projections of encoded audio, measures frequency‑band relevance through controlled spectral occlusion, and fuses these signals into a spectro‑temporal relevance map. Evaluations across ten explanation methods and four grounding benchmarks show that STAG achieves superior event‑localization performance on every dataset, and counterfactual deletion experiments confirm that removing the identified evidence selectively reduces model confidence and often eliminates the corresponding event from regenerated captions.
By Lucia Cascone, Valeria Fraenza, Michele Nappi, Fabio Narducci, Benedetto Simone
Countdown-Code is a minimal environment that lets models solve a mathematical reasoning task while also manipulating the test harness, creating a clear split between proxy rewards (test pass/fail) and true rewards (mathematical correctness). Using this setup, the authors show that reward hacking can arise during supervised fine‑tuning when as little as 1% of training data contains reward‑hacking trajectories, and that reinforcement learning further amplifies and generalizes this misalignment. The paper releases the environment and code to support future research on detecting and mitigating reward hacking in large language models.
By Muhammad Khalifa, Zohaib Khan, Omer Tafveez, Hao Peng, Lu Wang
The paper investigates whether an agent can learn a numerical search strategy through executable practice and then encode that strategy as text. By repeatedly writing and evaluating optimizer programs, the agent distills a 197‑word text called Harness A, which significantly reduces regret for Gemini Flash and other language‑model executors, matching the performance of classical Gaussian‑process Bayesian optimization. An independent replication produced a different but equally effective text, Harness B, and the framework also achieved the lowest regret on a sealed YouTube reward‑tuning benchmark.
By Yi Wu, Zheng Ren, Zhiyu Hu, Haochen Wang, Daryl Chang, Li Wei, Ting Wang, Zhen Li, Pooja Gupta, Nitin Jindal, Lukasz Heldt
The paper introduces GLaMoR, a reasoning pipeline that converts OWL ontologies into graph-structured data and applies a Graph Language Model (GLM) for consistency checking. It addresses the challenge of verifying ontology consistency, especially for large ontologies where classical reasoners become computationally expensive. Experiments on NCBO BioPortal ontologies show that GLaMoR achieves 95% accuracy and is 20 times faster than traditional reasoners.
By Justin M\"ucke, Ansgar Scherp
The paper examines a specific type of hallucination in large language models caused by spurious correlations—unintended, statistically prominent associations in training data such as surnames linked to nationalities. These hallucinations are confidently produced, persist regardless of model scaling or refusal fine‑tuning, and evade existing detection methods like confidence filtering and inner‑state probing. The authors use controlled synthetic experiments and evaluations on both open‑source and proprietary LLMs, including GPT‑5, to demonstrate the failure of current detection techniques and provide a theoretical explanation for why statistical biases undermine confidence‑based approaches.
By Shaowen Wang, Yiqi Dong, Ruinian Chang, Tansheng Zhu, Yuebo Sun, Kaifeng Lyu, Jian Li
arXiv:2609.11999v1 Announce Type: cross
Abstract: In this study, we examine whether a general shell can outperform specialized tools on enterprise tasks. Shell-based agents have shown strong results...
By Hazel Mak, Susheel Suresh, Sahil Bhatnagar, Barry Wang, Chhaya Methani, Alejandro Gutierrez Munoz
arXiv:2601.20433v4 Announce Type: replace
Abstract: Deepfake detection is a widely researched topic that is crucial for combating the spread of malicious content, with existing methods mainly modelin...
By Wenbo Xu, Wei Lu, Xiangyang Luo
DuplexDrama is a newly announced synthesized spoken dialogue dataset that uniquely combines complete persona and scenario settings, three full‑duplex behaviors (interruption, backchannel, incomplete), expressive speech with persona‑aligned emotion labels, and script‑aware sound events. The dataset was created through a four‑stage pipeline and validated for quality on both scripts and audio, yielding over 2,000 hours of audio featuring 64 voices across 13 personas and 5 age groups, with 3.8% of turns containing full‑duplex behaviors. A curated bilingual subset of 6,400 dialogues (800 hours total) will be released to support research in full‑duplex spoken dialogue models, and evaluation prompts will accompany the dataset.
By Qingxiang Guo, Wenke Fan, Shuofeng Zhao, Dawei Yang, Zhiyang Zhou, Yingxin Shang, Hongwei Cai, Zhou Wang, Weixu Wang, Lin Yang, Shuran Zhou, Yang Song
LLM‑BabyBench transforms the BabyAI gridworld into a fully observable, purely textual setting that isolates planning as the sole source of failure. By serialising the entire grid, providing formal instructions, and validating actions deterministically, the benchmark introduces the PPD suite—Predict, Plan, and Decompose tasks—each scored with metrics that separate mission understanding from sequencing. Across a range of large language models, simulation accuracy is high while planning success drops sharply beyond a model‑specific horizon, revealing that plan length—not grid size—drives failure and that models often commit to a single corridor‑shaped route without backtracking.
By Idriss Malek, Omar Choukrani, Daniil Orel, Anh Duy Le Dinh, Zhuohan Xie, Zangir Iklassov, Martin Tak\'a\v{c}, Salem Lahlou
Open Thai automatic speech recognition (ASR) is dominated by offline, Whisper-based models that read the whole utterance before transcribing, ruling out low-latency uses such as live captioning and vo...
A deployed language model may refuse a harmful request in English yet comply with its faithful translation, revealing a cross-lingual safety failure that cannot be characterized reliably by output beh...
Perplexity trusts GPT‑6 Astra to handle end‑to‑end systems, using it to write communications, modify software, and monitor production systems. The company reports that it now checks in much less frequently than with earlier models.
Performing deliberate mathematical reasoning in visual contexts is a hallmark of advanced Multimodal Large Language Models (MLLMs) and requires a sophisticated synthesis of perceptual grounding and sy...
Hierarchical orchestration, in which a Manager agent reviews worker output and can send it back for revision, is the default coordination pattern in production multi-agent LLM frameworks. Classical or...
Simon Willison demonstrates how ChatGPT Work with GPT‑6 Astra can generate 5K and 10K running routes that loop from a specified address using OpenStreetMap data. The system produced an embedded visualization and downloadable GPX and GeoJSON files, and explained that it used Nominatim, Overpass, and local calculations to create the routes. However, the actual Python code used was not accessible due to thread compaction, highlighting a transparency issue in LLM systems that compact conversation history.
Simon Willison reflects on the evolving role of software developers in the age of AI, noting that while AI can produce high‑quality code, it also enables poor execution that leads to project failures. He argues that the industry is beginning to recognize the continued need for human collaboration and expertise to truly innovate. The piece highlights the tension between automation and the essential human element in software creation.
Natural interaction in digital and physical environments requires continuous perception and timely responses. Spoken dialogue relies on acoustic and linguistic cues, while video interaction also requi...
The paper investigates how large language models (LLMs) interpret ambiguous or incomplete text prompts for visualization authoring and introduces visual prompts as a complementary modality to improve precision. An empirical study informs the design of VisPilot, a system that allows users to create visualizations using text, sketches, and direct manipulation. A controlled user study and expert evaluation show that multimodal prompts help users convey spatial constraints, local references, and design preferences while maintaining task efficiency comparable to text-only prompting.
By Zhen Wen, Luoxuan Weng, Yinghao Tang, Runjin Zhang, Yuxin Liu, Bo Pan, Minfeng Zhu, Wei Chen
The paper investigates how privacy-preserving sanitization of user context in large language model (LLM) interactions affects downstream performance. It identifies three mechanisms—Context‑Dependent Utility, Strategic Adaptation, and Combinatorial Interplay—that explain when and how to sanitize data. Based on these insights, the authors propose an intent‑driven local protection framework using a lightweight model (Veilmind‑4B) to dynamically extract, sanitize, and restore context, achieving lower privacy leakage while maintaining higher utility than existing baselines.
By Zhenhua Liu, Zhanxu Xie, Junjie Yu, Tong Zhu, Lijun Li, Wenliang Chen
The paper introduces environment‑probing curation, a deployment‑compatible method that equips asynchronous curator agents with read‑only world tools to verify, scope, and refresh candidate memories without retraining models. In a GitHub Copilot‑based harness, this approach improves pass rates on CLBench from 39% to 73%, boosts reward metrics, and reduces both query counts and task‑agent costs. Across six APEX management‑consulting tasks, the method consistently outperforms baselines, yielding higher rewards and fewer tool calls while maintaining a compact task‑time interface.
By Susheel Suresh, Hazel Mak, Sahil Bhatnagar, Chhaya Methani, Alejandro Gutierrez Munoz