Token Cap Placement Alters Test-Time Math Reasoning
New research measures how enforcing a token budget mid-derivation changes test-time mathematical reasoning outcomes.
Research Digest
Preprints and papers, rewritten in plain English.
Field experiment tests how AI voice agents sway children
A new randomized field experiment moves research on children's susceptibility to AI persuasion out of the lab and into real-world settings.
When Reasoning Models Get Stuck in Loops
A new study highlights how large reasoning models can fall into repetitive verification and generation cycles, undermining their efficiency.
New Metric Audits Whether LLM Summaries Preserve Employee Voice
A proposed metric quantifies how faithfully LLM-generated summaries of employee feedback represent the original voices, addressing a hidden layer of bias.
Benchmarking Voice Interfaces for LLM Computer-Use Agents
A new benchmark, Talk2Agent, highlights the gap between written and spoken instructions for LLM-based computer-use agents.
OpenJev-RLCD offers a working RLCD implementation for decision models
A new arXiv paper describes a working RLCD implementation, contrasting with the usual supervised fine-tuning and temperature scaling approach.
Code to Control: Generating Reactive Controllers to Cut LLM Latency
A new method synthesizes parameterized reactive controllers as code, avoiding per-decision LLM calls and online planning to reduce latency in real-time control.
Action-Conditioned Bisimulation Improves GUI Agent Memory
A new approach distinguishes web pages by what an agent does on them, not just how they look.
GraphCert Bootstraps Graph Agents with Evidence Rubrics
A new arXiv paper introduces GraphCert, a method that bootstraps agentic reasoning over knowledge graphs using certified evidence rubrics.
Population-curated training prevents stagnation in LLM algorithm-design agents
A proposed training method uses population-level curation to keep LLM-based algorithm-design agents from stalling during in-context evolution.
Classifier Reliability Can Drift Even When Confidence Distribution Holds Steady
New research quantifies the worst-case shift in a classifier's calibration under covariate changes that keep the confidence distribution fixed.
Diversity Combining Aims to Fix Plateau in Multi-Path LLM Reasoning
A new arXiv preprint argues that diversity combining can overcome the plateau that limits self-consistency in multi-path LLM reasoning.
AI Agent Swarms Show Promise as Autonomous Researchers
A new arXiv paper explores how swarms of coding agents can carry out multi-step research tasks with minimal human supervision.
PreferenceFlow Guides Robot Policies With Human Interventions
A new arXiv preprint proposes a test-time guidance method that uses human interventions to correct flow-matching robot policies without retraining.
StructRL targets long-horizon robot tasks with online RL
A new method applies online reinforcement learning to help vision-language-action models handle multi-step manipulation commands.
S2T-Unet Framework for Inter-Modality MRI Translation
A new structure-to-style framework aims to synthesize missing MRI modalities from existing scans, potentially reducing additional scanning.
DispFlow-GS Improves Dynamic Scene Reconstruction with Motion Disentangling
A new method enhances deformable 3D Gaussian splatting for monocular dynamic scenes by supervising displacement flow and separating motion components.
Erasing Violent Motions from Video Diffusion Models
New research targets unsafe actions like kicking and stabbing in text-to-video diffusion models through concept unlearning.