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 new approach, Loki, adapts pretrained models to predict new classes without additional training by using a metric that relates labels via distances. This can improve model performance on zero-shot prediction tasks.

This paper presents a predictive law for calculating the uplift of Large Language Model (LLM) ensemble performance based on diversity of thought. It provides an experimentally verified formal law and a compact heuristic for calculating uplift, which is tested on various datasets.

Researchers propose a benchmark (NARCBench) and probing techniques for detecting collusion between AI agents in multi-agent systems, achieving high detection rates in various scenarios.

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.

This paper proposes a hybrid deep learning framework for few-shot malware detection using an Autoencoder Feature Extractor and Model Agnostic Meta Learning. The model demonstrates high accuracy and robustness in adapting to limited data scenarios, but is not directly related to AI agents or their infrastructure.

Researchers propose a new method for asynchronous reinforcement learning in large language models, addressing the issue of stale rollouts generated by earlier policies. Their method, GMC-GRPO, provides improved convergence guarantees and better performance in experiments.

Researchers proposed an Online MARL framework via one-step Flow model (OMAF) that combines expressive generative policies with efficient one-step action generation, reducing training overhead and improving sample efficiency in online multi-agent settings.

Researchers investigate whether large language models (LLMs) can automatically formalize symbolic constraints for Neuro-Symbolic predictors, making them suitable for high-stakes applications. They introduce a benchmark to evaluate this and find that LLMs can generate formulas similar to human experts, leading to high-quality predictions.

The paper presents LOOM, a method for scaling looped mixture-of-experts (MoE) large language models (LLMs). LOOM stabilizes recurrence and diversifies computation across loops, enabling stable scaling to 9-12 loops. Experiments show improved performance in perplexity and zero-shot accuracy.

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.

GUI-HARVEST is an automatic harness optimizer for GUI agents with frozen backbone models. It aligns model outputs with observed action effects, accounts for execution variability, and identifies recurring failure patterns to improve agent performance by 12.33-13.87 percentage points on various benchmarks.

Researchers propose a post-training method for multi-token prediction heads in language models, achieving similar speedup to joint pre-training with significantly fewer tokens. They also introduce a relaxation of draft token verification and an adaptive controller for dynamic MTP head engagement.

A new method for selecting synthetic data to fine-tune large language models (LLMs) is proposed, focusing on maximizing the value of synthetic data to the target task. The method, called Training-Aware Target Coverage (TATC), is based on a linear theory that characterizes the tradeoff between the benefits and errors of synthetic data. Experimental results show that TATC outperforms alternative methods on various tasks.

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.

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.