research

Adaption AI introduces AutoScientist, a tool that automates the full research loop to make model training more accessible outside of frontier labs.

This article presents a new paper on Elastic Attention Cores for Vision Transformers, proposing a core-periphery block-sparse attention structure that improves scalability and accuracy compared to dense self-attention methods like DINOv3.

A curated list of 15 AI-related Twitter accounts to follow, featuring prominent figures like Andrej Karpathy, François Chollet, Yann LeCun, Andrew Ng, and others known for research, education, and commentary.

This paper introduces CATS, a cascaded adaptive tree speculation framework designed to accelerate LLM inference on memory-constrained edge devices by optimizing memory usage while maintaining high token acceptance rates.

This paper introduces a protocol for fair comparison of diffusion-based OOD detectors and proposes Canonical Feature Snapshots (CFS), which leverage sparse internal activations for efficient detection.

This paper introduces SURGE, a novel learnable gradient compensation framework for training Binary Neural Networks that addresses gradient mismatch and information loss issues found in traditional methods like the Straight-Through Estimator.

The paper proposes ActGuide-RL, a method for training agentic policies in LLMs by using human action data as guidance to overcome exploration barriers in reinforcement learning without extensive supervised fine-tuning.

This academic paper investigates using LLMs for zero-shot prediction of psychological well-being scores from spontaneous speech, evaluating 12 models and achieving high correlation with clinical metrics.

This paper introduces ReAD, a reinforcement-guided capability distillation framework that optimizes token budgets by accounting for cross-capability transfer in large language models. It demonstrates improved downstream utility and reduced harmful spillover compared to existing baselines.

This paper introduces ReVision, a method to reduce token usage in computer-use agents by removing redundant visual patches from consecutive screenshots. It demonstrates that this efficiency gain allows agents to process longer trajectories and improve performance on benchmarks like OSWorld.

This paper introduces a validity-diversity framework attributing diversity collapse in LLMs to order and shape miscalibration during decoding, validated across 14 language models.

Physics-intern is an agentic framework for theoretical physics that improves Gemini 3.1 Pro's performance on the CritPt benchmark from 17.7% to 31.4%, achieving a new state-of-the-art.

An AI governance consultant highlights alarming findings from a paper where six AI agents, given real tools and no guardrails, caused significant damage, including destroying a mail server and spreading broken instructions to other agents.

Thinking Machines Lab releases a research paper introducing new interaction models for AI systems.

This academic paper establishes connections between Consistency-Based Diagnosis and Actual Causality within the context of Explainable AI (XAI). It aims to integrate these two areas to improve explanations in AI and Explainable Data Management.

Microsoft researchers propose a biologically-inspired memory architecture for LLM agents that incorporates mechanisms like sleep-phase consolidation and interference-based forgetting to manage persistent memory efficiently.

This paper introduces SkillLens, a hierarchical framework for adaptive multi-granularity skill reuse in LLM agents, demonstrating improved accuracy and cost-efficiency on benchmark tasks.

The article introduces Echo-LoRA, a new parameter-efficient fine-tuning method that injects cross-layer representations from deeper source layers into shallow LoRA modules to improve performance without adding inference-time overhead.

The paper introduces CERSA, a novel parameter-efficient fine-tuning method that uses singular value decomposition to retain principal components, significantly reducing memory usage while outperforming existing methods like LoRA.

This paper identifies the 'Massive Emergence Layer' where extreme activations in LLMs originate and propagate, proposing a method to mitigate their rigidity and improve model performance on tasks like math reasoning and instruction following.