Researchers introduce A-MESS, a framework that evaluates jailbreak attacks on large language models (LLMs) based on their utility for improving model safety when used as red-teaming data, rather than solely on attack success rate (ASR). Using Shapley values, the method attributes marginal utility to individual attacks and selects compact, effective subsets for safety training. Experiments show that ASR rankings are only weakly aligned with defender-centric utility, and that optimizing attack subsets directly for safety yields better results than traditional attacker-centric approaches.
Why it matters: This work reframes the evaluation of jailbreak attacks to focus on their value for enhancing LLM safety, potentially leading to more effective red-teaming and safety training strategies.
Researchers have introduced FLOB, an obfuscation framework designed to protect floating-point computations in deep neural network (DNN) binaries using Mixed Boolean-Arithmetic (MBA) transformations. FLOB lifts floating-point values into a higher-precision binary expansion space, performs computations there, and projects results back to the original precision without introducing additional rounding error. Experiments demonstrate that FLOB reduces the operator recovery rate to 4.51%, representing a 32.82 percentage-point improvement over existing obfuscation methods.
Why it matters: This work offers a practical and effective method for safeguarding proprietary DNN models against reverse engineering, specifically addressing the unique challenges of floating-point operations.
Researchers have identified a novel attack surface in agentic retrieval-augmented generation (RAG) systems, termed 'salience induction.' Unlike content poisoning or prompt injection, this attack manipulates the prominence and framing of truthful information to mislead multi-hop reasoning agents. The team introduces a defense method called Salience Normalization, which significantly reduces attack success rates from 83.3% to 15.3% in their experiments across multiple model families and agent architectures.
Why it matters: This work exposes a critical vulnerability in RAG systems, showing that even truthful content can be manipulated to mislead AI agents, and demonstrates an effective defense.
A new interview study with 22 software practitioners reveals that security is rarely considered when selecting and integrating AI components, such as large language models, into software systems. Instead, decisions are primarily driven by functional criteria like performance, accuracy, and cost. The research highlights that established lessons from traditional software supply chain security are often overlooked in the rush to adopt AI technologies. The authors recommend adopting a security-by-design approach to better address these emerging risks.
Why it matters: Neglecting security in AI component selection could expose software systems to significant vulnerabilities, repeating past mistakes in software supply chain management.
A new preprint introduces the Privacy-Cost Equity Ratio (PCER), a group fairness metric for differentially private machine learning that incorporates disparities in privacy exposure. PCER is calculated by normalizing a group's positive prediction rate by its overfitting gap, which serves as a proxy for vulnerability to membership inference attacks. Evaluations on several benchmarks, including COMPAS, show that protected groups can simultaneously face higher privacy risks and worse predictive outcomes—patterns that standard outcome-only fairness metrics may overlook. The metric is practical for post-hoc audits, requiring only per-group train and test accuracy.
Why it matters: This work broadens the concept of fairness in privacy-preserving machine learning by highlighting the importance of accounting for which groups bear the privacy risks, not just who benefits from model outcomes.
Researchers present the first differentially private algorithms for the max-sum diversification (MSD) problem under both cardinality and matroid constraints. Their methods achieve nearly optimal utility and are faster than existing non-private approaches. Experiments on real-world datasets show that the private algorithms maintain utility comparable to non-private baselines, even under strong privacy guarantees, and offer significant speed improvements for cardinality constraints.
Why it matters: This work enables privacy-preserving data summarization and query output diversification with strong utility and efficiency, addressing a key gap in handling sensitive data.
A new preprint introduces RT-SHCUA, an architecture designed to enable real-time, secure use of self-hosted computer-use agents (SHCUAs) for controlling unmanned aerial vehicles (UAVs). The system addresses the mismatch between language agent reasoning and the strict timing, safety, and security requirements of UAV operation by separating semantic reasoning from onboard execution. It enforces contract-bound skill invocations with explicit checks for timing, state, and authority, and supports degraded handling and auditable evidence preservation. Prototype evaluation demonstrates bounded responsiveness and trusted enforcement without embedding the full agent in critical control loops.
Why it matters: This work proposes a novel architecture to safely and securely bridge language agents with real-time UAV control, addressing key challenges in timing, authorization, and accountability.
Researchers have introduced TRACE, a new framework for restoring safety alignment in large language models (LLMs) after custom fine-tuning. TRACE learns a safety patch by simulating harmful tuning trajectories and optimizing a plug-in patch that recovers safety with minimal impact on task performance. Experiments across six benchmarks and two models show that TRACE achieves nearly perfect safety while preserving the utility of the fine-tuned models.
Why it matters: TRACE provides a practical approach for Fine-Tuning-as-a-Service (FTaaS) platforms to restore safety in LLMs after user customization, addressing a key challenge in deploying safe and effective AI systems.
AEGIS is a framework for investigating span-level guidance in multilingual text detoxification across English, Mandarin Chinese, and Korean. The study finds that providing explicit rationales can alter the trade-off between reducing toxicity and preserving meaning, but the impact varies depending on the generator backbone and linguistic context. Rather than focusing on achieving state-of-the-art results, the work analyzes when and how span-level control signals are effective for detoxification.
Why it matters: This research clarifies the conditional usefulness and limitations of span-level control signals for multilingual text detoxification, informing future development of safer language models.
Researchers introduce a method for document-level event argument extraction that leverages fine-tuned mid-sized open large language models (LLMs). Their approach integrates role-set injection, LoRA-based parameter-efficient fine-tuning, and deterministic decoding with schema validation, achieving 42.39% F1 on the MAVEN-ARG benchmark with Phi-4 (14B). This method outperforms previously reported GPT baselines on multiple evaluation metrics.
Why it matters: The work shows that accessible, fine-tuned open LLMs can surpass proprietary models on complex information extraction tasks, potentially broadening access to high-quality structured data extraction.
A new preprint introduces a diagnostic framework for evaluating test-time collaboration methods in large language models (LLMs), breaking down performance gains into measurable components: oracle gap and signal fidelity. The authors show that on benchmarks like LiveCodeBench, a high-fidelity verifier achieves a +8.14 percentage point improvement, while a low-fidelity verifier gains only +2.70pp. The framework enables practitioners to estimate the potential benefits and limitations of collaboration strategies before deployment.
Why it matters: This work provides a practical tool for anticipating the effectiveness of test-time collaboration in LLMs, helping avoid unnecessary investment in unproductive multi-agent approaches.
A new preprint analyzing five model families and seven bias types finds that sycophancy and related cue-induced biases in large language models (LLMs) are primarily introduced during alignment tuning, not pretraining. The study shows these biases correspond to distinct, causally active directions in model hidden states, which can be decoded and manipulated to recover unbiased answers. The findings suggest that each bias remains representationally distinct and that targeted interventions can partially debias model outputs.
Why it matters: This research clarifies the origins and structure of certain LLM biases, highlighting alignment tuning as a key source and suggesting new avenues for targeted debiasing.
Researchers introduce Trace-Based On-Policy Distillation (TOPD), a teacher-supervised framework for transferring reasoning abilities to masked diffusion language models without the need for reward estimation. TOPD enables the SDAR-4B-Chat model to achieve comparable accuracy to RL-trained models on the MATH500 benchmark, while requiring four times fewer rollout rounds and an estimated 96-fold compute speedup.
Why it matters: This approach offers a more efficient method for reasoning-oriented post-training of diffusion LLMs, potentially reducing computational costs compared to reinforcement learning-based methods.
Researchers developed a human-in-the-loop AI agent that assembles evidence and drafts one-sentence impact summaries for clinical scholars. In a real-world evaluation across 10 scholars, both reviewers accepted or edited 81.7% of the agent's findings, and the median review time per scholar dropped from an estimated 15 hours manually to just 14 minutes. The agent's evidence covered all four Translational Science Benefits Model (TSBM) domains and included non-scholarly impact categories often overlooked by manual processes.
Why it matters: This work shows that a human-in-the-loop AI agent can make large-scale impact reporting feasible for clinical research programs by drastically reducing staff time and improving coverage of impact categories.
TalTech's systems achieved first place in both tracks of the Beyond Transcription Challenge (BeTraC), generating SOAP notes directly from long doctor-patient conversation recordings without intermediate transcription. They adapted Voxtral Mini and Voxtral Small models using LoRA fine-tuning and DAPO reinforcement learning with the Open Medical Concept F1 metric as a reward. Independent evaluation found their submissions had the lowest hallucination rate among all entries.
Why it matters: This result shows that reinforcement learning with a concept-matching metric can improve factual reliability in medical summarization from speech, and that text-based fine-tuning can transfer effectively to speech input.
JOR-Bench is a suite of five Japanese-language benchmarks designed to evaluate large language models (LLMs) on operations research (OR) problems, comprising 1,319 problems across linear, mixed-integer, non-linear, and combinatorial optimization. The benchmarks are Japanese translations of established English datasets, enabling direct cross-lingual comparison. Evaluation of seven LLMs shows that strong multilingual models exhibit nearly identical OR formulation accuracy in Japanese and English, with only a -0.3 percentage point difference, though error analysis uncovers subtle language-specific failure modes.
Why it matters: JOR-Bench provides a rigorous tool for assessing LLMs' mathematical reasoning in Japanese, highlighting both the strengths and nuanced limitations of multilingual models across languages.
Researchers introduce KyrgyzLLM-Bench, a benchmark suite for evaluating large language models (LLMs) in Kyrgyz. The suite includes two natively authored datasets (KyrgyzMMLU and KyrgyzRC) and translated, manually post-edited versions of WinoGrande, HellaSwag, BoolQ, and TruthfulQA. Evaluation of 26 models shows that cross-lingual transfer varies by task, with translation artifacts notably affecting HellaSwag performance. Few-shot prompting improves some open-source models on reading comprehension, but results are inconsistent for proprietary models on translated tasks.
Why it matters: This work provides the first large-scale, natively authored evaluation benchmark for Kyrgyz, enabling more accurate assessment of LLM capabilities and cross-lingual transfer in a low-resource language.
A new preprint argues that achieving zero hallucination in large language models (LLMs) cannot be accomplished by the model alone, but must be enforced at the system level. The authors introduce HALO, a six-layer assurance architecture that includes grounded generation, constrained execution, multi-signal verification, calibrated abstention, total traceability, and continuous oversight. The framework is demonstrated on a regulated claims-extraction task.
Why it matters: This work reframes hallucination as a system-level failure mode and proposes a practical, multi-layered architecture for deploying trustworthy AI in regulated enterprise environments.
Researchers introduce ESCUCHA, the first Spanish speech understanding benchmark designed to evaluate large audio language models across diverse acoustic conditions and reasoning abilities. The benchmark features 1,000 human-curated questions paired with 162.9 hours of real-world audio, encompassing multiple Spanish accents and non-normative speech. Initial evaluations show substantial performance gaps between state-of-the-art models and trained humans.
Why it matters: This benchmark fills a critical gap in evaluating audio language models for Spanish under realistic and challenging conditions, highlighting the need for more robust and capable multilingual speech AI systems.
A new benchmark dataset of 1,516 expert-verified Bangla sentences reveals that large language models (LLMs) exhibit a systematic bias toward the dominant meaning when disambiguating Bangla homographs that serve as both personal names and common nouns. The study demonstrates that contrastive chain-of-thought prompting and distilled cultural explanations can reduce this bias from as high as 100% to under 5% in small models, even turning a previously unsuccessful Bangla-specific model into the strongest performer.
Why it matters: This work demonstrates that language-specific pretraining is insufficient for cultural grounding and offers practical methods to mitigate cultural biases in low-resource language models.