The Morning
From the arXiv
Open-MMUnlearning: Unifying Methods and Evaluation for MLLM Unlearning
his paper introduces Open-MMUnlearning, a unified, open-source framework for evaluating machine unlearning in multimodal large language models (MLLMs). It addresses the fragmentation in existing methods by providing standardized interfaces for preparing models, processing data, applying unlearning techniques, and evaluating performance across privacy, safety, and copyright concerns. The framework's key contribution is enabling systematic and comprehensive assessment of MLLM unlearning by supporting diverse models, methods, and benchmarks.
Before They Can Solve: Predicting Post-Training Coding-Agent Performance from Base Models
This paper proposes a method to predict the potential of base models for agentic coding tasks without expensive post-training. It identifies the "decisive step" within a successful agent trajectory – the first code-changing action that makes the repository pas…
Fault-tolerant foundation models
This paper demonstrates that large language models (LLMs) can be trained to be resilient to unreliable hardware, with error tolerance increasing with model size. The core method involves modifying neural scaling laws based on training on simulated faulty hardw…
MIRA: A Musical Intent Refinement Agent for Aligning Text-to-Music Generation with User Intent
This paper addresses the challenge of aligning text-to-music generation with user intent by proposing a rubric-based evaluation framework. Their core method, MIRA (Musical Intent Refinement Agent), refines prompts for black-box text-to-music models by iterativ…
PHRBench: A Behavioral Evaluation of Post-Hallucination Reasoning in LLMs
This paper introduces PHRBench, a novel benchmark designed to evaluate how Large Language Models (LLMs) handle and correct hallucinated information during multi-stage reasoning. The core method involves a controlled evaluation of LLM reasoning trajectories, ca…
Reasoning-Token Spikes Under Prompted Untruthful Responding in Large Language Models
This paper proposes a new method for detecting untruthful responses in large language models by monitoring the *number* of reasoning tokens generated, rather than their content. Th…
RECAST: Learning to Compute the Right Context through Adaptive Evidence Routing
RECAST addresses the limitation of traditional RAG by learning to actively *derive* evidence rather than just retrieve it. Its core method involves a RouterLM that sequentially dec…
RoboQuest: Generalist Physical Agents that Search, Inspect and Test
RoboQuest introduces a benchmark for embodied agents that must actively explore and gather information through physical interaction to complete tasks in unfamiliar environments. It…
SemanticFold: Latent Sequence Compression SeparatesLanguage Modeling, Decodability, and Reasoning
SemanticFold is a novel method that compresses prompt prefixes in large language models by folding hidden states at learned boundaries. This technique aims to determine if such com…
SLDR: Defending Against Malicious Fine-tuning via Selective Layers Recovery and Dynamic Routing
SLDR defends against malicious LLM fine-tuning by selectively recovering layers most sensitive to safety. It trains a lightweight adapter on these critical layers and uses dynamic …
The Town Square
Meta and Microsoft are implementing measures to limit their employees' use of Claude AI, likely due to concerns about data security and intellectual property.
Workshops
This repository provides a curated collection of "skills" or reusable code snippets, inspired by the author's personal `.agents` directory, designed to empower real engineers with practical tools.
AnyPS5 is a tool that automates the process of porting PlayStation 5 executables to Linux and Windows.
