Voice Agents Keep Live Channel, Delegate Heavy Tasks Async
A new arXiv paper describes a converging design where voice agents stay on a full-duplex live channel while offloading search, reasoning, and coding to asynchronous delegation.
Research Digest
Preprints and papers, rewritten in plain English.
Mine Odyssey Benchmark Tests Agents' Spatial Intelligence
A new arXiv paper introduces Mine Odyssey, a benchmark for evaluating how well AI agents explore and understand unfamiliar physical environments.
SynCo: Co-Training Data Synthesis for Self-Evolving LLMs
A new arXiv paper proposes SynCo, a multi-agent reinforcement learning approach to help LLMs self-evolve with less manual supervision.
How Perturbations Propagate in Diffusion Model Sampling
A new study examines why equal-sized modifications to diffusion sampling can produce very different changes in generated outputs.
Defensive Writing: How GPT Models Retract Authors' Claims
A new arXiv preprint identifies a pattern where ChatGPT revisions narrow or retract claims that the source material does not support.
Rethinking Environmental Feedback in Agentic Hindsight Self-Distillation
A new arXiv paper argues that the way environmental feedback is modeled is critical for training language agents when rewards are unavailable.
New Benchmark Tests Calibration and Framing Sensitivity of System One Models
TypedBench provides a standardized way to evaluate decision models that output probabilities over typed answers, focusing on calibration, framing sensitivity, and cost.
AI Emotion Recognition vs. Affective Science: A Process View
A new arXiv paper contrasts categorical emotion recognition with affective science's view of emotion as an unfolding process.
EvoSim preprint: building physics models as a learnable process
A new arXiv preprint frames the laborious craft of building physics-based models as a task that can itself be learned.
Calibrating Reasoning Models via Reasoning-Guided Uncertainty
A new arXiv paper proposes RL-ARC, which uses reasoning-guided uncertainty to calibrate large reasoning models trained with reinforcement learning.
Book Argues Artificial General Intelligence Is Mathematically Impossible
A new book-length argument contends that AGI cannot be achieved because intelligence arises from complex human processes that resist mathematical formalisation.
Benchmark's 'Selective Forgetting' Scores May Just Reflect Last-Write-Wins
A new analysis suggests MemoryAgentBench's conflict-resolution results can be reproduced by a frozen rule that always takes the newest statement, undercutting the 'selective forgetting' interpretation.
Audits Incentivize Unsafe-Task Reporting but Training May Not Follow
A study finds that audits can make reporting unsafe tasks optimal in multi-agent systems, but do not ensure teams learn to report.
Sigma-Hunter: A Specialized LLM for Threat Detection Rules
A new domain-specific language model aims to help detection engineers translate threat reports into valid, testable Sigma rules.
BEACON-SP: Ontology-Grounded GraphRAG for Suicide Risk Assessment
A new framework applies ontology-grounded graph retrieval-augmented generation to support clinicians in assessing suicide risk.
PAIR Method Targets the Perception-to-Action Gap in VLA Models
A new arXiv paper proposes an approach called PAIR to help vision-language-action models convert scene descriptions into robot actions.
Heterogeneous GNNs for Shared-Roadmap Generation in Multi-Agent Path Planning
A new arXiv paper proposes using heterogeneous graph neural networks to generate shared roadmaps for multi-agent path planning, aiming to improve safety and search efficiency over traditional methods.
Taming High-Recall LLM Outputs with Dataset-Adaptive Post-Processing
A new arXiv paper proposes dataset-adaptive post-processing to turn messy, high-recall LLM extractions into clean, useful customer intents.