AI Theory

Thinking Intervention offers a novel paradigm for controlling reasoning in LLMs, enabling fine-grained guidance and improvements in instruction-following and safety.

Model Merging: An unlearning system, which combines specialized models, achieves top results in SemEval-2025 Task 4 by selectively erasing sensitive knowledge.

LookAhead Tuning: Safer LLMs via Partial Answer Previews by preserving initial token distributions.

LLMs’ reasoning is decoded via sparse autoencoders, revealing key features that, when steered, enhance performance. First mechanistic account of reasoning in LLMs!

AI progress is tracked with a new metric, 50%-task-completion time horizon, showing exponential growth with a doubling time of ~7 months, hinting at significant automation potential in the near future…

Multi-Agent Systems (MAS) often underperform despite enthusiasm. This paper analyzes 5 popular frameworks across 150+ tasks, identifying 14 failure modes categorized into specification/design, inter-a…

AI reasoning models reveal bias-like patterns, processing association-incompatible info with more computational effort, mirroring human implicit biases.

Group-robust machine unlearning via MIU reduces perf. degradation in dominant groups after unlearning, preserving model robustness without compromising accuracy.

MoE-X: An intrinsically interpretable Mixture-of-Experts language model that uses sparse, wide networks to enhance transparency.

CodeCriticBench: A new benchmark for holistic code critique by Large Language Models.

AI system access is more than just release; it’s about how accessible system components are, impacting benefits, risks, and scalability.

Predicting open-weight AI model growth on Hugging Face using a citation-style model, revealing adoption dynamics and influencing factors.

LLMs use punctuation in context memory, surprisingly boosting performance by using seemingly trivial tokens.

LLMs fail to act safely when considering user-specific safety standards, which were made to be solved via new benchmark.

RL optimizes quantum error-correcting codes, slashing physical qubit overhead for fault-tolerant quantum computing.

Aligned LLMs’ safety often anchors in the template region, creating vulnerabilities. Detaching safety mechanisms shows promise in mitigation.

Unveiling the Brownian snake within the Brownian sphere! This research constructs the inverse of the CVS bijection, mapping the sphere back to its underlying snake.

LLMs personalize based on user names, but this study reveals that cultural presumptions in LLM responses risk reinforcing stereotypes.

ASTRAL, a novel automated safety testing tool, reveals DeepSeek-R1’s significantly higher unsafe response rate compared to OpenAI’s o3-mini, highlighting critical safety concerns in advanced LLMs.

Researchers used ASTRAL to systematically test OpenAI’s 03-mini LLM’s safety, revealing key vulnerabilities and highlighting the need for continuous, robust safety mechanisms in large language models.