Eitan Sprejer

Aligning LLM-based judges with human preferences is a significant challenge, as they are difficult to calibrate and often suffer from rubric sensitivity, bias, and instability. Overcoming this challenge advances key applications, such as creating reliable reward models for Reinforcement Learning from Human Feedback (RLHF) and building effective routing systems that select the best-suited model for a given user query. In this work, we propose a framework for modeling diverse, persona-based preferences by learning to aggregate outputs from multiple rubric-conditioned judges. We investigate the performance of this approach against naive baselines and assess its robustness through case studies on both human and LLM-judges biases. Our primary contributions include a persona-based method for synthesizing preference labels at scale and two distinct implementations of our aggregator: Generalized Additive Model (GAM) and a Multi-Layer Perceptron (MLP).

Chain-of-thought (CoT) outputs let us read a model's step-by-step reasoning. Since any long, serial reasoning process must pass through this textual trace, the quality of the CoT is a direct window into what the model is thinking. This visibility could help us spot unsafe or misaligned behavior (monitorability), but only if the CoT is transparent about its internal reasoning (faithfulness). Fully measuring faithfulness is difficult, so researchers often focus on examining the CoT in cases where the model changes its answer after adding a cue to the input. This proxy finds some instances of unfaithfulness but loses information when the model maintains its answer, and does not investigate aspects of reasoning not tied to the cue. We extend these results to a more holistic sense of monitorability by introducing verbosity: whether the CoT lists every factor needed to solve the task. We combine faithfulness and verbosity into a single monitorability score that shows how well the CoT serves as the model's external 'working memory', a property that many safety schemes based on CoT monitoring depend on.

The core premise of AI debate as a scalable oversight technique is that it is harder to lie convincingly than to refute a lie, enabling the judge to identify the correct position. Yet, existing debate experiments have relied on datasets with ground truth, where lying is reduced to defending an incorrect proposition. This overlooks a subjective dimension: lying also requires the belief that the claim defended is false. In this work, we apply debate to subjective questions and explicitly measure large language models' prior beliefs before experiments. Debaters were asked to select their preferred position, then presented with a judge persona deliberately designed to conflict with their identified priors. Our main findings show that models tend to prefer defending stances aligned with the judge persona rather than their prior beliefs, sequential debate introduces significant bias favoring the second debater, models are more persuasive when defending positions aligned with their prior beliefs, and paradoxically, arguments misaligned with prior beliefs are rated as higher quality in pairwise comparison.

Feature steering has emerged as a method for controlling LLM behavior through direct manipulation of internal representations, offering advantages over prompt engineering. However, its practical effectiveness in real-world applications remains poorly understood, particularly regarding potential trade-offs with output quality. We show that feature steering methods substantially degrade model performance even when successfully controlling target behaviors. We evaluate Goodfire's Auto Steer against prompt engineering baselines across 14 steering queries on 171 MMLU questions using Llama-8B and Llama-70B, measuring accuracy, coherence, and behavioral control. Our findings show that Auto Steer successfully modifies target behaviors but causes dramatic performance degradation: accuracy drops from 66% to 46% on Llama-8B and 87% to 73% on Llama-70B, with coherence falling from 4.62 to 2.24 and 4.94 to 3.89 respectively. These findings highlight limitations of current feature steering methods for practical deployment where task performance cannot be sacrificed.