Feedback Isn't Always Helpful for Learned Scientific Controllers
A new audit finds a learned controller can beat a misspecified planner even when it gains nothing from feedback.
Research Digest
Preprints and papers, rewritten in plain English.
MemDev-Bench: Benchmarking Memory Systems for Code Agents
A new benchmark evaluates whether memory systems help code agents reuse experience across real repository histories.
Structured-Token Attention Proposed for mmWave Liquid Identification
A new attention-based approach targets fine-grained classification of similar liquids using compact millimeter-wave measurements.
Runtime Control for AI Agents in Operational Workflows
A new arXiv paper proposes runtime control to close the gap between evaluation and execution for language-model agents handling approvals and commitments.
Benchmarking Whether LLM Agents Stay Within User Authorization
New research proposes a benchmark to test whether LLM agents act beyond the scope users authorize.
Assessing Question Value in On-Policy Distillation
A new arXiv paper frames training questions as carriers of distributional knowledge and asks how to judge the value of newly generated ones.
New Framework Quantifies LLM Influence on Trust and Decisions
A proposed behavioral computing framework aims to measure how subtle linguistic cues in LLM outputs affect user trust and decision-making.
TriGeoBench tests multimodal LLMs on geometry with figures and TikZ
A new benchmark aims to expose how well multimodal large language models reason about geometry, going beyond rasterized images.
Meta Expert shares MoE experts across layers to cut storage and retrieval costs
A proposed Mixture-of-Experts design shares experts across layers to make model scaling more storage- and retrieval-efficient.
Multi-agent AI collaboration may hit a capability ceiling
A systematic study of language-model agents in dual-control settings suggests that the benefit of adding more agents may diminish as individual agent capability increases.
SCODEL: A Shared Ontology for Recipe Entity Linking
A new framework aims to resolve ambiguous and context-dependent food mentions in recipes by grounding them in a shared canonical ontology.
Pairwise CNNs Infer Cognitive Traits From Limited Brain Scans
A new arXiv preprint proposes a pairwise convolutional neural network approach to link limited neuroimaging data to continuous cognitive traits.
AI Model Uncovers Hidden Patterns in Healthcare Claim Underpayments
A new unsupervised framework uses LLM interpretation to surface systematic payer underpayment and denial patterns at hospital scale.
E8-QAT: 2-Bit Quantization Training with Implicit Lattice Codebooks
A new arXiv paper proposes an efficient 2-bit quantization-aware training method using implicit E8 lattice codebooks to reduce LLM memory footprint while preserving accuracy.
Reusing Multimodal Agents to Evaluate Game Sound Designs
A new framework uses multimodal learning and reinforcement learning to make in-game sound evaluation more efficient.
Why Experience Memory Gains in LLM Driving Agents May Be Unreliable
A controlled evaluation argues that claimed benefits of experience memory in closed-loop LLM driving agents are confounded by warmup pool composition and prompt-token effects.
Slot-JEPA: Object-structured prediction for latent world models
A new arXiv preprint argues that replacing dense patch representations with object-structured slots can give latent world models a cleaner prediction interface.
Knowledge-Guided Evolution for LLM-Generated Ascend NPU Kernels
A new paper argues that LLM agents, though strong on GPUs, need knowledge-guided evolution to generate Ascend NPU kernels where pretraining data is scarce.