Cross-Interface Training Boosts LLM Agent Robustness
A simple consistency-training method helps LLM agents stay reliable when interfaces change without altering the task.
Research Digest
Preprints and papers, rewritten in plain English.
Finite safety tests can't exclude adaptive AI adversaries
A new paper shows that finite safety assessments of embodied and tool-use AI cannot rule out adversaries that attack only after certification.
Agentic Scientist Framework Automates Computational Research
A new arXiv preprint describes a system that uses coding agents to write code, run in silico experiments, and analyze results with minimal human input.
Zero-Shot Language Models for Name-Based Record Linkage
A new paper shows pre-trained language models can link records and infer demographics from names without labeled training data.
How Language Models Shift From Copying to Composition in In-Context Learning
A new arXiv study investigates how in-context learning develops, moving beyond accuracy to examine how contextual information controls predictions.
Attention-guided perturbation reduces video LLM hallucinations
A new method strengthens contrastive decoding for video language models by using attention maps to perturb representations, cutting unsupported responses.
Plan or Retrieve First? New Study Probes Tool Discovery in LLM Agents
A scale study examines whether LLM agents should plan before retrieving tool definitions or retrieve first, given finite context windows.
Semi-Amortized Inference Reuses Local Posteriors for Hard Bayesian Models
A new arXiv paper proposes a semi-amortized strategy that reuses local posteriors to make Bayesian inference tractable in high-dimensional models.
AI Framework Targets Timing-Aware Power Recovery After Routing
An AI-driven framework is proposed to decide cell changes in post-route power recovery by accounting for the cost of repairing side effects.
Adaptive Gaussian Mixture Model Detects Wireless Spoofing at Physical Layer
A proposed adaptive Gaussian mixture model aims to detect physical-layer spoofing attacks that manipulate channel state information in wireless localization.
TorchSpin brings differentiable simulation to EPR spectroscopy
A new PyTorch-based simulator aims to make electron paramagnetic resonance spectrum fitting more efficient through differentiability.
Echo cuts tensor-parallel communication to 4 bits with error compensation
A new method called Echo reduces the all-reduce traffic in tensor-parallel LLM training and inference to 4-bit values while compensating for the resulting error.
SANZARU benchmark tests AI teams under information asymmetry
A new benchmark shows that an AI model's individual skill does not guarantee success when it must cooperate with uneven information and communication.
Active Learning Targets Rare Document Classes
A new arXiv paper explores active learning to reduce the cost of labeling rare documents like targeted hate speech.
SAFE-KV repairs causal state for non-prefix KV cache reuse in RAG
A new arXiv preprint proposes a suffix-repair method that lets retrieval-augmented generation reuse cached document computations even when the query context does not share a prefix.
New Benchmark Tracks AI Agents' Sustained Progress on Long Tasks
A new benchmark evaluates AI agents on whether they can maintain progress through extended workflows, not just reach a correct final state.
Correction Training Aims to Fix Diffusion Models' Exposure Bias
A new training approach targets the mismatch between training and inference inputs that causes error accumulation in diffusion models.
UDHRBench: A New Benchmark for Human Rights in LLMs
The benchmark evaluates large language models against the Universal Declaration of Human Rights.