Longer Thinking May Make Reasoning Models Take 'Some' Literally
A new arXiv paper asks whether giving reasoning models more time to think changes how they handle everyday scalar inferences like 'some' and 'or'.
Research Digest
Preprints and papers, rewritten in plain English.
Information Density in Reasoning Steps May Predict LLM Correctness
A new arXiv study explores whether the density of information in each chain-of-thought step can predict and steer a language model's answer correctness.
SPRA Gives Scientific Agents a Structured View of Papers
A new paper-level representation aims to help AI agents connect problems, methods, and findings scattered across narrative research articles.
Do Syntax-Focused Attention Heads Actually Drive Model Predictions?
A new study asks whether attention heads that track syntactic relations are causally responsible for a language model's grammatical predictions.
Choosing Attention Heads by Intrinsic Dimension
A new heuristic estimates the intrinsic dimension of token representations to set the number of attention heads in advance, avoiding grid search.
AI Search Summaries May Cut Traffic to Content Platforms
A new arXiv paper examines how generative AI answers on search pages could disrupt the referral traffic that sustains platforms like Reddit.
AI Value Alignment Should Be Personal, Not Global
A new arXiv preprint argues that language model alignment needs to account for individual and contextual values rather than applying one global standard.
OrionMoE: A Mixture-of-Experts Model for Tabular In-Context Learning
A new arXiv preprint proposes a mixture-of-experts architecture for tabular foundation models, aiming to improve in-context learning without task-specific updates.
Diagnostic Test Probes LLM Chemical Predictions Beyond Accuracy
A new arXiv paper argues that standard chemical prediction benchmarks may let LLMs exploit target identity information, and proposes a diagnostic test to reveal whether models actually understand chemistry.
Input ambiguity for LLMs is model-dependent, new paper argues
A new framework suggests that whether a prompt needs clarification depends on the language model's own set of possible interpretations, not just the input text.
VASPO: Value-Guided Self-Distillation for Reasoning Models
A new method corrects a flaw in on-policy self-distillation by using privileged value estimates to align supervision with expected return.
Study Tests Whether Gemini Brand Mentions Hold Up After 20 Days
A pre-registered replication repeats the same requests to Gemini Answers 20 days later to see if brand-mention counts are stable enough to act on.
Measuring the Limits of What Language Models Can Articulate
A new arXiv paper proposes a framework for quantifying the upper bounds of what LLMs can express through language.
Plug-in decision heads could replace text generation for web agents
A new arXiv paper argues that web agents waste tokens by generating full action strings when a simple decision head would do.
RCCM Diagnoses When Reasoning Prompts Help or Hurt Deployments
A new diagnostic framework explains why the same reasoning prompt can improve one deployment while degrading another.
Selective World Modeling Proposed for Sparse-Reward Language Agents
An arXiv paper suggests that language agents should model only decision-relevant parts of the world to cope with sparse trajectory-level rewards.
Attributing AI Text to the Model That Wrote It
A new arXiv paper proposes a robust method for identifying which large language model generated a given piece of text, addressing misuse in academia and beyond.
Hindsight-Privileged RLVR Offers Token Credit Without Group Sampling
A new reinforcement learning method aims to solve the credit-assignment problem for individual reasoning tokens in verifiable-reward tasks.