This paper explores a novel method of inducing vulnerabilities in large language models (LLMs) using 'drunk language', which can lead to jailbreaking and privacy leaks. The researchers found that LLMs are more susceptible to these vulnerabilities than previously reported approaches.

VISPA is a training-free framework for pluralistic alignment of large language models, enabling direct control over value expression by dynamic selection and internal model activation steering. It achieves performant results across various pluralistic alignment modes in healthcare and beyond, and is adaptable with different steering initiations, models, and/or values.

Researchers propose UniGuardian, a training-free detector for Large Language Models (LLMs) that identifies prompt injection, backdoor, and adversarial attacks without knowing the attack type. UniGuardian measures how prompt perturbations shift the model's output distribution and uses a single-forward strategy for efficient detection and text generation.

A benchmarking framework, AstroAgentBench, is introduced for evaluating agentic planning in space mission planning tasks. It assesses the performance of large language model (LLM) agents in domains like scheduling, observation planning, and constellation design. The framework provides a standardized evaluation methodology and highlights the importance of task-contract formulation, verifier feedback, and search adaptation in achieving high-quality plans.

KaliBench is a fine-grained benchmark for evaluating the ability of language models to generate executable commands for real-world cybersecurity tools. It includes 8,504 query-command pairs across 1,642 tools and enables precise and reproducible assessment of tool selection and argument construction.

SourceLearn, a new learning mechanism for developing reusable source-specific competence in large language model agents, improves understanding of persistent authoritative sources by representing this competence with a persistent source model and using two complementary learning mechanisms: Self-Directed Source Learning and Task-Guided Source Learning.

This paper proposes Transferable Example Scoring and Selection (TESS), a scalable data-selection framework for training large language models. TESS uses a Pointwise Value Matching objective to improve transferability and generalization.

A new method called CARM (Cancellation-Aware Response Masking) is proposed to address off-policy issues in large language model (LLM) reinforcement learning. CARM improves upon existing sequence-level masking by taking the absolute value of token log-ratios, preventing policy drift and resulting in better performance on mathematical reasoning and code generation tasks.

Mem++ is a non-destructive memory framework for long-term organizational LLM agents. It stores documents whole with their date and author, allowing for read-time selection of relevant documents and fusion of lexical and semantic rankings. Evaluations show Mem++ outperforms baseline memory systems by 8-13 points and achieves the best overall score on the gpt-4.1-mini model.

MoLE is a new framework for latent visual reasoning in vision-language models, which encourages different latent tokens to extract complementary visual information. It outperforms existing methods on five visual reasoning benchmarks, achieving an average score of 78.6.

Researchers introduced VeriSpec, a novel approach to detect inconsistencies in model specifications using large language models as verifiers. VeriSpec extracts rules from the specification text, clusters related rules, and applies LLM-as-verifier reasoning to identify inconsistencies. The approach achieved high precision and efficiency in detecting defects in the OpenAI Model Spec, outperforming several baselines.

Researchers propose a new approach to mitigating model collapse in iterative fine-tuning using a non-parametric entropy rate estimator. The approach does not require a model or external data and shows significant improvements in text diversity metrics.

Paper explores the effectiveness of cross-model review in LLM verification, finding that a second model does not always improve error detection.

Researchers introduced JOINGR, a join-aware table retrieval method for Text-to-SQL, treating the database join graph as the retrieval space. It selects semantically similar anchor tables, traverses join edges, and aggregates scores. JOINGR improves recall over baselines on the BEAVER benchmark and transfers across benchmarks.

Researchers introduce Task Operator (TO), a method to improve in-context learning (ICL) efficiency by analytically deriving updates to attention output projection. TO achieves better performance than prior methods and enables many-shot scaling without expanding the context window.

Researchers investigated how weak reviewers can audit strong coding agents. They found that official execution evidence can improve defect catch and reduce over-rejection. The study used 411 execution-labeled traces from three agents and 101 controlled cases.

Researchers propose an alternative reconstruction loss function (RMSE) to improve post-training quantization of large language models (LLMs) by decoupling optimization strength from the reconstruction loss scale.

Researchers propose a method called 'grafting' to apply pre-training interventions to AI models without requiring a full post-training run, reducing iteration time and improving model stability.

A research paper proposes a new method for compressing large language model attentions, called FTC, which improves perplexity on several downstream tasks without requiring fine-tuning or gradient-based recovery.

Researchers studied the relationship between LLM iterates and discovery success, developed 12 harnesses called 'Modular', and found that initialization improves LLM-driven discovery. Their results suggest that early discoveries are predictive of eventual success and propose a universally applicable intervention for initialization.