This research explores security vulnerabilities in autonomous AI coding agents, demonstrating how persistent-state environments allow agents to distribute malicious payloads across multiple pull requests.
An empirical study investigates whether the theoretical computational advantages of Mixture-of-Experts models translate to faster and cheaper inference on consumer and edge hardware.
Microsoft Research has introduced Flint, an open-source visualization language designed to help AI agents generate expressive charts from compact specifications.
This paper explores the transition of large language model-driven theorem provers from solving well-defined problems to assisting with frontier mathematical research and open conjectures.
A new compete-then-collaborate framework ranks frontier AI models to build a verifiable curriculum for training smaller student models in coding tasks.
Researchers have found that the step-by-step reasoning processes in advanced language models introduce vulnerabilities that attackers can exploit to slow down systems.
Researchers have introduced ReCoLoRA, a framework that uses spectrum-aware recursive consolidation to enable continual fine-tuning of large language models without losing performance on prior tasks.
Researchers introduced Latent Personality Alignment, a method that aligns language model safety using a small set of harm-agnostic statements instead of large datasets of harmful prompts.
Researchers developed video-SALMONN-R3, a framework that improves video large language models by first localizing relevant segments and then re-analyzing them at higher resolutions.
This study investigates the cognitive gap in large language models where newly memorized facts during fine-tuning fail to generalize to downstream reasoning tasks.
ResonatorLM is proposed as an alternative language modeling architecture that uses causal resonant field mixing to improve long-context processing efficiency.
Researchers investigate whether training policies to predict both actions and environmental changes using World Action Models improves robot manipulation learning from egocentric human videos.
Peer-Predictive Self-Training is a collaborative, label-free fine-tuning framework where multiple language models use aggregated responses to mutually improve their reasoning capabilities.
A new study demonstrates that consistency and agreement among language models are unreliable indicators of accuracy when evaluating AI outputs.
A new subword tokenizer called Thunder-Tok reduces token fertility and inference costs for large language models while maintaining downstream performance.
The OpenCoF framework introduces a method for models to perform logical reasoning through temporally connected video frames, termed Chain-of-Frame reasoning.
This study analyzes how explainable artificial intelligence methods can facilitate the safety certification and regulatory compliance of machine learning models.
This paper addresses feature entanglement in language models by proposing a method to encourage almost orthogonal features, enabling more precise and isolated mechanistic interventions.
The paper presents MultiSim, an ensemble approach that utilizes multiple driving simulators to identify and mitigate testing discrepancies in automated driving assistance systems.
A new auditing technique called overthinking uses task vectors to amplify a model's reasoning process, helping to expose hidden information or misaligned behaviors.
Jet-Long introduces a dynamic bifocal rotary position embedding method to efficiently extend the context window of large language models in a zero-shot manner.
A new study reveals that training models on data following a power-law distribution, rather than a curated uniform distribution, actually improves their ability to perform compositional reasoning tasks.
Researchers have demonstrated that hackers can exploit security vulnerabilities in popular large language models to construct large-scale botnets through a technique called HalluSquatting.
Researchers have introduced ParamMute, a method that improves the faithfulness of retrieval-augmented generation by suppressing knowledge-critical feed-forward networks that cause models to ignore retrieved context.