A new arXiv preprint finds that combining information from multiple anonymized speech utterances—using audio, prosody, and text—significantly improves the ability to verify a speaker's identity, even when anonymization techniques are applied. The study shows that multimodal systems outperform unimodal ones, and that aggregating as few as five anonymized utterances can reduce error rates by over 15% compared to audio-only methods. This suggests that current anonymization methods may not fully protect speaker privacy when adversaries have access to multiple utterances and multimodal data.
Why it matters: The findings highlight a significant privacy risk, indicating that widely used speaker anonymization techniques may be vulnerable to re-identification attacks using multimodal aggregation.
A new arXiv preprint introduces OpenSkillRisk, a benchmark designed to assess how well large language model (LLM) agents avoid unsafe actions when using third-party skills. Testing 13 LLMs and three agent frameworks on 263 real-world risky skills, the study finds that even the safest systems still execute unsafe actions in about 17% of cases. The analysis identifies recurring failure modes, including not recognizing risks, failing to intervene, and over-following instructions.
Why it matters: The results highlight significant unresolved safety risks for LLM agents that rely on third-party skills, raising concerns for real-world deployment.
A new arXiv preprint formally defines 'structural generalization' and uses computational complexity theory to argue that pure Transformer architectures cannot learn this property under standard assumptions (TC0 ≠ NC1). The authors further contend that neuro-symbolic systems only succeed at structural generalization by hard-coding part of the solution, and that current benchmarks do not distinguish between genuinely learned and pre-specified rules.
Why it matters: This work challenges the prevailing assumption that Transformers can achieve structural generalization, raising fundamental questions about the capabilities and evaluation of modern neural language models.
A recent arXiv preprint introduces a taxonomy of six types of 'silent failures' in multimodal agentic search systems, where errors in the reasoning process are masked by correct final answers. The authors develop a diagnostic pipeline to evaluate both answer correctness and evidence-grounding quality, finding that standard surface accuracy metrics can significantly overestimate true system reliability across several leading multimodal models.
Why it matters: The work highlights that widely-used evaluation methods may overlook critical reliability issues in advanced AI systems, underscoring the need for more thorough diagnostics to ensure trustworthy deployment.
A new arXiv preprint argues that calibration, the standard method for evaluating confidence in large language models (LLMs), is insufficient because it allows for incoherent and unfaithful probability estimates. The authors introduce a new framework with three axes—structural coherence, faithfulness, and usefulness—to more rigorously assess LLM uncertainty. They find that commonly used confidence estimators can appear well-calibrated while still violating these coherence criteria, indicating that current LLM confidence scores may not represent true probabilistic beliefs.
Why it matters: This challenges the reliability of LLM confidence estimates, raising concerns about their trustworthiness in applications where accurate uncertainty quantification is critical.
A new arXiv preprint introduces ITPEval, a benchmark designed to evaluate automated translation of formal proofs between four widely used interactive theorem provers: Lean 4, Rocq, Isabelle, and HOL Light. The benchmark includes over 1,500 source files and nearly 7,000 theorems, and assesses both statement and proof translation using several large language models. Results show that proof translation remains challenging, with a maximum pass@1 rate of 10.5%, and that mismatches between libraries are a major obstacle.
Why it matters: This work provides the first large-scale, multi-system benchmark for proof translation, highlighting key challenges for interoperability and data sharing in formal mathematics and AI-driven theorem proving.
A new arXiv preprint introduces Learn2Discern (L2D), a benchmark designed to test large language models' (LLMs) ability to weigh information from external sources. Evaluating 13 models across nearly 670,000 trials, the study finds that LLMs perform near chance at distinguishing reliable sources and updating beliefs toward the truth. While newer and larger models show some improvement in truth discernment, they do not improve at recognizing source reliability, highlighting a persistent limitation.
Why it matters: This finding raises concerns about the reliability of LLMs as they are increasingly used to access and evaluate information online.
METR researchers coauthored a paper analyzing how AI might accelerate its own R&D through feedback effects, sometimes referred to as recursive self-improvement (RSI). The paper decomposes these feedback effects and highlights uncertainty about whether AI capabilities growth will accelerate or plateau due to various bottlenecks. It also clarifies the different definitions of RSI and focuses on the strength of feedback for forecasting future capabilities.
Why it matters: This analysis informs forecasts of AI capabilities growth, which is important for assessing future AI risk.
A new preprint provides a tight characterization of the types of inexact score oracle access that allow for sampling with vanishing total variation bias in a standard target family. The main result demonstrates that any error weaker than the sub-Gaussian assumption precludes tractable unbiased sampling, extending previous work to be algorithm-agnostic and applicable to broader error models.
Why it matters: This result clarifies the theoretical limits of sampling with imperfect score estimates, which is important for understanding the reliability of score-based generative models such as diffusion models.
Researchers propose Posterior Prefix Tuning (PPT), a method for steering the behavior of Bayes-filtered transformers (BFTs) by optimizing prompts using importance sampling from the latent posterior. PPT does not require forward passes or backpropagation through the transformer, and a single set of prior samples can be reused for multiple utility functions at minimal additional cost. The method is validated on Beta-Bernoulli and reinforced urn BFTs across three types of utility functions.
Why it matters: PPT provides a computationally efficient alternative to gradient-based optimization for adapting transformer models to diverse objectives.
Researchers introduce PDDIM, a new algorithm for solving linear inverse problems using diffusion priors. The method modifies standard DDIM updates with coordinate-wise adjustments based on signal-to-noise ratio, and is proven to converge to the Bayesian posterior. Empirical results demonstrate that PDDIM performs favorably compared to existing diffusion-based posterior samplers across various image restoration tasks.
Why it matters: This work offers a practical and theoretically grounded approach to posterior sampling in inverse problems, combining empirical effectiveness with provable guarantees.
Researchers propose partial fusion of neural networks, a method that interpolates between full ensembles and weight aggregation to balance computational cost and model performance. The technique aggregates only the most similar neurons across networks, using partial optimal transport for matching, and generalizes to a form of pruning that allows for isolating, deleting, or combining neurons. This approach provides a flexible framework for managing the tradeoff between accuracy and efficiency in neural network deployment.
Why it matters: This method could make neural network ensembles more practical by reducing their computational demands while preserving much of their performance.
Researchers have introduced QuArch, the first benchmark specifically designed to evaluate large language models' (LLMs) knowledge and reasoning abilities in computer architecture. QuArch includes 2,671 expert-validated question-answer pairs spanning topics such as processor design and memory systems. Leading LLMs achieved between 34% and 73% accuracy on advanced questions, and fine-tuning on QuArch led to improved performance on a realistic memory hierarchy design task, yielding more area-efficient and viable hardware solutions.
Why it matters: QuArch addresses a significant gap in LLM evaluation by providing a domain-specific benchmark for computer architecture, enabling more rigorous assessment and development of AI capabilities in this critical field.
Researchers introduce a discrete denoising diffusion framework that leverages a sample-efficient estimator (NeurISE) for single-site conditional probabilities, eliminating the need for score function approximations. Experiments on binary datasets, including Ising models, MNIST, and quantum annealer data, show that this method outperforms existing approaches in several evaluation metrics.
Why it matters: This work presents a more sample-efficient approach to discrete diffusion, potentially advancing generative modeling for scientific and binary data.
Researchers propose latent distribution matching (LDM) as a new theoretical framework that unifies various self-supervised learning (SSL) methods, including contrastive, non-contrastive, and predictive approaches. The framework also leads to a nonlinear Bayesian filtering model for high-dimensional time series and demonstrates identifiability of latent representations under mild conditions. LDM clarifies the assumptions behind existing SSL methods and offers principled guidance for developing new approaches.
Why it matters: This work provides a unified theoretical foundation for self-supervised learning, potentially enabling more systematic and effective development of representation learning methods.
A new controlled longitudinal study investigates the impact of agent harness evolution on coding agent quality, finding that quality regressions often attributed to large language models (LLMs) are actually caused by updates to the agent harness. By evaluating 35 sequential releases of the Qwen Code CLI while keeping the underlying LLM constant, the study reveals that quality fluctuations are linked to specific development patterns and architectural changes in the harness.
Why it matters: This research challenges the common assumption that LLMs are primarily responsible for coding agent regressions, highlighting the significant influence of agent harness middleware and the importance of its careful management.
Researchers introduce SWE-Milestone, a benchmark designed to evaluate AI agents on streams of milestone-level coding tasks that simulate real-world software evolution. By testing 12 advanced models across 4 agent frameworks, they observe that performance drops from over 80% on isolated tasks to 38.03% in continuous, evolving scenarios. This highlights significant challenges for AI agents in maintaining system integrity and managing error propagation over time.
Why it matters: This benchmark reveals a major limitation in current AI coding agents: their inability to reliably handle long-term software maintenance, which is essential for real-world applications.
TraceDev is a multi-agent framework designed to convert natural language requirements into repository-level code using five specialized agents and a traceability graph. In evaluations on two datasets, TraceDev achieved success rates of 53.63% and 56.82%, outperforming state-of-the-art baselines by up to 340.80%. The framework emphasizes explicit requirement traceability, supporting more reliable code generation for complex software scenarios.
Why it matters: TraceDev introduces a novel approach to requirement traceability in automated code generation, addressing a key limitation in current LLM-based software development methods.
Researchers present Spaghetti Architect, a tool that generates code datasets in five programming languages (Python, JavaScript, Go, Java, C++) with controlled difficulty and resistance to data contamination. The tool produces correct-by-construction programs from a known-optimal reference, allowing for reliable evaluation of code models. It enables fine-grained control over code 'messiness' and labels each instance along intrinsic and incidental difficulty axes.
Why it matters: This tool addresses the persistent challenge of data contamination in code model evaluation by enabling the creation of fresh, controlled, and contamination-resistant datasets.
Researchers introduce LISA, a framework that leverages large language models (LLMs) to generate API sequences and program invariants for detecting functional bugs in software libraries. LISA achieves higher bug-detection rates and competitive code coverage compared to traditional fuzzing and previous LLM-based test generation methods.
Why it matters: LISA offers a new approach to uncovering non-crashing functional bugs that are often missed by conventional fuzzing techniques, potentially improving software reliability.