The paper reports that large language models can acquire cipher-based covert communication skills without fine‑tuning, using prompting or in‑context learning instead. This enables new jailbreak attacks that bypass alignment safeguards by encrypting harmful requests, making them appear as nonsensical text to harmfulness classifiers. The authors demonstrate successful attacks against frontier models from Anthropic, Google, and OpenAI.
By Thomas Rivasseau
BRACE introduces an anchored Bellman‑residual correction to address stale critic bias in asynchronous reinforcement learning for language models. By limiting the correction horizon to a prefix of policy tokens and adding a constant‑weight Monte‑Carlo tail, it separates policy correction from reward propagation. The method improves mean@1 on BrowseComp‑Plus by 2.4% and runs 2.46× faster per step than synchronous training while staying stable 50 updates off‑policy.
By Guanqun Zhao, Zijun Xie, Binbin Zheng, Jiafeng Lu, Enlei Gong, Zeyu Chen
The paper titled "With a Thermomix You Lose the Ability to Cook: A Kitchen Machine Analogy for Applications of Generative AI in Education" uses the Thermomix kitchen appliance as an analogy to explore how generative AI tools like ChatGPT are adopted in education. By mapping Thermomix use cases onto learning scenarios and situating them within the ICAP and SAMR frameworks, the authors illustrate how different modes of tool use can either support or undermine meaningful engagement and learning. The study emphasizes that the key issue is not merely whether learners use AI, but how such use shapes their learning processes, offering a conceptual lens for researchers and practitioners to critically examine and guide AI integration in education.
By Nikol Rummel, Valentina Nachtigall, Ernesto Panadero
KairosAgent is an agentic framework that combines a large language model (LLM) reasoner with a time series foundation model (TSFM) forecaster to tackle cross‑domain multimodal time series forecasting. It dynamically invokes analytical tools to improve the LLM’s numerical comprehension and semantic reasoning, then fuses the reasoning outcomes into the TSFM pipeline for more accurate predictions. The approach is further enhanced by a curated large‑scale trajectory corpus and a reinforcement learning paradigm with multi‑turn refinement and turn‑level credit assignment, achieving superior zero‑shot forecasting performance.
By Kun Feng, Ziwei Shan, Yuchen Fang, Yiyang Tan, Sihan Lu, Shuqi Gu, Xingyu Lu, Lintao Ma, Kan Ren
The study investigates whether adding more components to a P300 brain‑computer interface speller always improves performance. Using a full‑factorial experiment on a public dataset, the authors varied Euclidean Alignment, xDAWN spatial filtering, subject calibration, and language model priors, and found that component effects are conditional rather than additive. Calibration was the most influential, while adding components could sometimes reduce performance, demonstrating component anti‑synergy and challenging the assumption that an ‘all‑on’ pipeline is best.
By Lucas Yang, Rui Liu, Fusheng Wang
PRAGMA is a benchmark designed to evaluate personalized guidance in long‑term conversations. It includes curated longitudinal conversation histories, evidence annotations, and guidance scenarios that reflect evolving user contexts and incorrect assumptions. Experiments show that current retrieval, memory, and long‑context models struggle to recover relevant conversational evidence and to use it effectively for personalized guidance.
By Hyojeong Yu, Hyukhun Koh, Minsung Kim, Yunah Jang, Kyomin Jung
The paper investigates how synthetic data used for fine‑tuning large language model (LLM) skill routers can lead to catastrophic forgetting of real and out‑of‑distribution (OOD) skills. By evaluating a production router with 34,396 skills, the authors show that synthetic‑data fine‑tuning improves in‑distribution retrieval but degrades performance on real and OOD data. They test several continual‑learning inspired mitigation methods—embedding‑anchor regularization, Learning without Forgetting, Elastic Weight Consolidation, and L2‑initialization—and find that these approaches both preserve OOD retrieval performance and boost synthetic in‑distribution retrieval by up to 13.98% for a 0.6B Qwen retriever and reranker.
By Syed Shariyar Murtaza, Yifan Nie, Utkarsh Soni, Eugene Wen, Arvid Frydenlund
The paper identifies a new failure mode in neurosymbolic systems called Verdict‑Preserving‑Unfaithfulness (VPU), where incorrect formal encodings can still pass solver checks. It introduces Generative Verification (GenV), a method that uses a language model to produce a continuous reference‑equivalence score without relying on explicit localization. Experiments show GenV+HN achieves high AUROC, generalizes to unseen translators, and improves downstream agent performance by 11.3 points.
By Vikash Singh, Debargha Ganguly, Aman Goel, Ali Torkamani, Xiaoxue Han, Joseph Lilien, Ferhat Erata, Vipin Chaudhary
DriftNet is a dual‑head trajectory Transformer designed to detect and localize prompt injection attacks in large language model agents. It processes logged tool‑call trajectories, classifying each as compromised or not while labeling every step as benign, injection point, hijacked, or failed injection. On the AgentDrift benchmark, DriftNet achieves high accuracy, with an F1 score of 0.983, 98.7% exact injection‑point recovery, and low false‑alarm rates.
By Asif Pinjari, Mithun Paul Saint-Germain
ZipCodec is a streaming neural speech codec that operates at an ultra‑low frame rate of 6.25 Hz and a bitrate of 0.80 kbps, achieving a theoretical latency of 160 ms. It leverages large‑scale WavLM distillation, a redesigned transformer architecture, a scalar spherical quantizer, and a latency‑aware streaming decoder to preserve reconstruction quality while reducing frame rate. Experiments demonstrate that ZipCodec outperforms existing streaming codecs at comparable bitrates in both reconstruction and downstream tasks, and it can run real‑time single‑stream inference on a consumer‑grade CPU despite having 842 M parameters.
By Luca Della Libera, Cem Subakan, Mirco Ravanelli
The paper introduces the Explainability Assistant, an open‑source conversational XAI system designed to interpret complex energy consumption forecasting models. By leveraging large language model function‑calling, it achieves 94% intent‑parsing accuracy and supports flexible natural‑language interaction across different ML problem types without task‑specific fine‑tuning. A comparative evaluation with energy domain specialists shows improved usability and consistent task accuracy, with all experts preferring the conversational interface over a traditional XAI dashboard.
By Rodion Krjut\v{s}kov, Eduard Barbu, Nikos Sakkas, Sofia Yfanti
The paper introduces Murmur2Vec, a lightweight, alignment‑free embedding that uses k‑mer counts hashed with MurmurHash to create a compact representation for biological sequences. It provides a full theoretical analysis, including bias/variance formulas, a Johnson–Lindenstrauss‑style concentration bound, and an excess‑risk bound that clarifies the trade‑off between hash‑table size and classifier performance. Empirically, Murmur2Vec matches or surpasses a fine‑tuned 650M‑parameter ESM‑2 protein language model across several classification tasks, including SARS‑CoV‑2 spike lineage and HIV‑1 Env subtype identification.
By Sarwan Ali, Taslim Murad, Imdadullah Khan, Safi Faizullah
The paper introduces SEM‑HD, a framework that leverages longitudinal mammography history as privileged information during training to improve risk prediction while requiring only a single current exam at inference. By having a student model predict latent representations of past visits and using teacher supervision from actual longitudinal data, SEM‑HD preserves temporal modeling benefits without needing prior exams at deployment. Experiments on three cohorts and two backbone architectures show consistent gains in long‑horizon AUC and pAUC, especially in low false‑positive‑rate regions, and recover much of the performance gap to full‑history models.
By Banafsheh Karimian, Soufiane Belharbi, Alexis Guichemerre, Luke McCaffrey, Mohammadhadi Shateri, Eric Granger
The paper introduces DP-Muon, a differentially private optimization method that incorporates matrix‑orthogonalized momentum. It employs standard global per‑example clipping and releases a single Gaussian‑noised gradient per step, with matrix and auxiliary updates treated as post‑processing. The authors analyze the mean distortion introduced when fresh Gaussian noise passes through a nonlinear matrix map, deriving exact Gaussian heat identities and showing that for a smooth Newton‑Schulz map, the conditional output bias is reduced from second to fourth order in the noise scale. They also establish matrix‑block stationarity bounds, quantify orthogonalization error, and provide criteria for improving the upper bound, while a separate inequality captures the impact of auxiliary Adam updates. Experiments on GPT‑2 at various privacy targets demonstrate that DP‑Muon configurations outperform Adam baselines in test negative log‑likelihood.
By Jihwan Kim, Chenglin Fan
The paper introduces a new perspective on entity rarity in multimodal entity linking by using knowledge‑graph structural metrics instead of popularity metrics, revealing many rare entities previously overlooked. Experiments show that state‑of‑the‑art models suffer a 15.4–39.9% accuracy drop on these rare‑entity slices. The authors propose a training‑free framework that combines reasoning and retrieval with a vision‑language model, achieving a 6.9% overall accuracy gain and up to 23.3% improvement on rare entities, and release a new benchmark MERLIN‑Rare for focused evaluation.
By Parinthapat Pengpun, Simran Khanuja, Graham Neubig
The paper investigates how transformer-based models and traditional feature-based models capture readability signals across five languages using the ReadMe++ dataset. By applying SHAP to identify key features for traditional classifiers and then probing XLM‑R and language‑specific encoders with TCAV, the authors find that transformers recover surface‑length, syntactic, and lexical‑diversity cues and reflect the CEFR ordinal structure, though alignment varies by model family, language, and layer. The study highlights that high linear separability does not guarantee directional influence, cautioning against overreliance on linear probing for readability features.
By Joshua Wong, Chris Tanner
The paper proposes using semantic uncertainty, derived from large language models, to predict Transition Relevance Places (TRPs) in spoken dialogue. By sampling possible continuations of an ongoing turn and measuring changes in semantic dispersion, the authors identify moments when a listener might take the floor. Their method outperforms prompt-based and fine-tuned text-only baselines on a dataset with real-time TRP labels, supporting the idea that evolving semantic constraints inform turn‑taking opportunities in unscripted interaction.
By Muhammad Umair, Jan P. de Ruiter
The paper argues that verbalized confidence—once viewed as overconfident and coarse—has become the preferred soft‑scoring method for LLM‑as‑a‑Judge on top‑tier proprietary models released after 2025. Experiments on SummEval, AggreFact, and HelpSteer2 across up to 18 LLMs show that log‑probabilities are no longer the best signal, and that adding an overconfidence advisory and self‑debate further improves calibration and robustness. The authors note that these enhancements incur little accuracy loss on post‑2025 models but do affect pre‑2025 ones, highlighting a compatibility shift in how confidence should be measured.
By Yu-Chung Hsiao
ProMediConv is a new benchmarking framework for evaluating proactive conversational agents in legal dispute mediation. It models mediation as a multi-stage, party-aware dialogue that incorporates 11 mediation strategies and four party behavior pattern states, and it is built on 972 real-world cases with utterance-level annotations. The framework introduces a fine-grained metric, MAD (Mean Attribute Difference), to capture shifts in party behavior throughout the dialogue, and provides a comprehensive benchmark with diverse models and a tailored baseline, ProMediAgent.
By Zesheng Wei, Mengfan Li, Wenhao Liu, Yixin Zhang, Zilei Wang, Yang Deng
FlexComp is a framework that allows a single model to perform context compression at any desired ratio, unlike existing methods that require separate models for each fixed ratio. It achieves this by sampling a memory budget during training and selecting the appropriate budget at inference time using either confidence-based cascade routing or a lightweight learned predictor. Experiments on ICAE, 500xCompressor, and SAC show that FlexComp matches the performance of specialized fixed-ratio models while enabling high compression rates and improving decoding throughput.
By Kaiyan Zhao, Zhongtao Miao, Akiko Aizawa, Yoshimasa Tsuruoka