Scaling-Down Laws Aim to Take Guesswork Out of LLM Compression
A new framework models how compressed language models lose capability, potentially turning empirical compression choices into predictable trade-offs.
Research Digest
Preprints and papers, rewritten in plain English.
Time-to-Recidivism Disparities Show Fairness Beyond Algorithms
A new preprint argues that racial disparities in recidivism persist in time-to-event data, suggesting fairness work must extend beyond algorithmic tweaks.
Detecting 'Mirages' in Vision-Language Models
A new preprint focuses on identifying when vision-language models give confident answers that lack supporting visual evidence.
UniFLM: A Unified AI Model for Fetal Limb Ultrasound Analysis
A new AI framework jointly performs segmentation and measurement on fetal limb ultrasound images, targeting lethal skeletal dysplasias that existing models often miss.
Google AI Overviews Reach 2 Billion Users, Study Finds
A new arXiv paper examines the scale, transparency, and quality of Google's AI-generated search answers.
SurGe refines local surface geometry in 3D point maps
A new method targets the persistent problem of inaccurate local surface geometry in feedforward 3D reconstruction.
Infinity Search: Approximate Vector Search via Ultrametric Projections
A new arXiv preprint proposes using ultrametric spaces to accelerate approximate vector search, though the abstract cuts off before detailing the results.
Long-Context LLMs Reduce Sequential Calls in Reranking
A new reranking method uses long-context language models to compare many documents at once, cutting the number of sequential model calls.
Unifying Differential Privacy Notions via Information Theory
A new framework treats four curve-based privacy notions as instances of a single information-theoretic model, potentially simplifying privacy accounting.
Modeling Future Wrist Views Improves Fine-Grained Robot Manipulation
A new approach treats wrist-view observations as task-conditioned future states rather than parallel inputs, aiming to improve fine-grained robot manipulation.
FlowReader Routes Evidence Budgets Across Long Multimodal Documents
A new framework treats evidence assembly as a min-cost flow problem, allocating a fixed budget across text, tables, figures, and slides while avoiding near-duplicates.
Digital Twin Maps ICS Telemetry to ATT&CK with Evidence Reasoning
A new framework uses digital twins and dependency reasoning to infer which ATT&CK techniques produced observed ICS process changes.
Testing Whether Audio-Language Models Ground Music Concepts in Sound
A new arXiv paper probes whether large audio-language models' musical understanding is tied to actual acoustic evidence.
Strengthening FROST2: New Security Enhancements for Threshold Schnorr Signatures
A new paper proposes practical security enhancements for FROST2, a threshold Schnorr signature scheme widely used in blockchain systems.
LiBRA: A New Attack That Removes Watermarks by Optimizing Latent Images
LiBRA is a detection-aware watermark removal method that uses bidirectional latent optimization to evade both watermark decoding and detection.
HeadEdit: Using the Frozen Unembedding Matrix to Fix LLM Behavior
A new approach targets persistent behavioral errors in language models by editing the unembedding matrix rather than treating them as computation problems.
New Bangla Dataset Targets Regional Dialects in Abusive Text
A balanced Bengali-script dataset aims to help NLP models handle regional dialect variation in informal and abusive Bangla text.
Jev: A Commercial System One Model That Returns Choices, Not Text
A new arXiv paper benchmarks Jev, a TypeSafe AI model that answers typed questions with fixed options, rubric scores, or probabilities rather than generating text.