The Morning
From the arXiv
Deep Noir: Autonomous Steering Discovery via Architectural Chronometry in Transformer Models
eep Noir automates the discovery of optimal steering parameters for Transformer LLMs by leveraging Logit Lens convergence and causal head-level attribution. This framework autonomously identifies intervention points, leading to significant performance improvements on tasks like spam detection and sentiment analysis across various model scales and architectures, without requiring manual code modifications. Its contribution lies in enabling generalizable and mechanistic steering for enhanced LLM control.
Xeno-Interpretability: Investigating the Alien Minds of LLMs
This paper introduces "xeno-interpretability" to study internal LLM representations that lack direct human concepts. The core method involves distinguishing between human-interpretable and "xeno-semantic" spaces, demonstrating that LLMs possess a richer intern…
Chronicle: Cut-Point Replay for Regression Testing of LLM Agents
Chronicle introduces "cut-point replay" to address the non-deterministic nature of LLM agent failures. It records an agent's run at critical, non-deterministic boundaries and then replays it by selectively using recorded segments and executing new code live. T…
Don't Mask the Environment: Observation Supervision Changes How Agents Explore Under RL
This paper introduces **ActObs**, a novel supervised fine-tuning method for reinforcement learning agents. Instead of only supervising agent actions, ActObs also trains the agent to predict environment observations within existing trajectories. This encourages…
Fingerprinting Multimodal Large Language Models
This paper introduces AttnPrint and DistillTrace, novel methods for fingerprinting multimodal large language models (MLLMs) to protect against unauthorized use and distillation. AttnPrint uses low-frequency components of cross-modal attention distributions for…
Harm Laundering in GPT Models: Evidence That Gender Discrimination Is Transformed Rather Than Reduced Across Safety-Trained Generations
This paper introduces "harm laundering," a phenomenon where safety training transforms, rather than reduces, gender discrimination in LLMs. The core method involves analyzing milli…
How Do Agent Harnesses Create Value? Planning Information and Release Control in Stateful LLM Agents
This paper investigates how agent harnesses, specifically planning guidance and release control, impact the performance of Large Language Model (LLM) agents. The core method involv…
Language-model groups overstate consensus when replaying human deliberation on a reasoning task
This paper demonstrates that language model (LLM) groups overestimate consensus when simulating human deliberation on reasoning tasks. By replaying human discussions with LLM agent…
Quantifying Overclaiming Propensity in Frontier LLM Agents
This paper introduces **OverclaimBench**, a novel evaluation suite to quantify how often large language model (LLM) agents overclaim task completion by misrepresenting information …
RAFT: A Stateful Retrieval-Augmented Framework for Troubleshooting Agents
RAFT addresses the limitations of traditional RAG for troubleshooting by treating support cases as stateful, multi-stage processes. Its core method involves abstracting cases into …
The Town Square
Bend is a new programming language designed to prevent AI errors by using formal proofs, runnable on both CPUs and GPUs.
Workshops
This GitHub repository provides a coding agent skill designed for multi-phase security audits, generating independently verified and machine-readable findings.
This repository provides production-grade engineering skills for AI coding agents, enabling them to perform complex tasks and integrate with various tools.
