arXiv subjects

Jeremias Ferrao

Publications and source records attributed to Jeremias Ferrao.

5 recordsLinked to original sources

arXiv2026

Paved with True Intents: Intent-Aware Training Improves LLM Safety Classification Across Training Regimes

We argue that safety classifiers should model user intent as an explicit signal between the prompt and the final label. To study this, we introduce AIMS, a human-annotated dataset of 1,724 difficult safety prompts, each paired with an intent description and harm label. We use AIMS to evaluate intent-aware training across supervised fine-tuning, preference learning, reasoning distillation, and reinforcement learning. Despite its size, AIMS enables competitive safety classifiers across training regimes: DPO from model-generated intent errors improves over SFT, and intent-conditioned distillation outperforms reasoning-only distillation in most teacher-student pairs. Most notably, directly rewarding intent faithfulness with GRPO yields the strongest average performance across five external safety benchmarks, while our intent-aware models form the inference latency-F1 Pareto frontier. These results show that faithful intent modeling is a compact, high-quality supervision signal for more robust safety classifiers.

cs.CL↗

arXiv2025

The Anatomy of Alignment: Decomposing Preference Optimization by Steering Sparse Features

Prevailing alignment methods induce opaque parameter changes, obscuring what models truly learn. To address this, we introduce Feature Steering with Reinforcement Learning (FSRL), a framework that trains a lightweight adapter to steer model behavior by modulating interpretable sparse features. First, we theoretically demonstrate that this mechanism is expressive enough to approximate the behavioral shifts of post-training processes. We then apply FSRL to preference optimization and perform a causal analysis of the learned policy. Our analysis reveals a crucial insight: the model learns to reward stylistic presentation as a proxy for quality, disproportionately relying on features related to style and formatting over those tied to alignment concepts like honesty. By effectively optimizing the preference objective, FSRL serves as a transparent proxy for observing the alignment process. Overall, FSRL offers an interpretable control interface and a practical way to diagnose how preference optimization pressures manifest at the feature level.

cs.AI↗

arXiv2025

What Really Counts? Examining Step and Token Level Attribution in Multilingual CoT Reasoning

This study investigates the attribution patterns underlying Chain-of-Thought (CoT) reasoning in multilingual LLMs. While prior works demonstrate the role of CoT prompting in improving task performance, there are concerns regarding the faithfulness and interpretability of the generated reasoning chains. To assess these properties across languages, we applied two complementary attribution methods--ContextCite for step-level attribution and Inseq for token-level attribution--to the Qwen2.5 1.5B-Instruct model using the MGSM benchmark. Our experimental results highlight key findings such as: (1) attribution scores excessively emphasize the final reasoning step, particularly in incorrect generations; (2) structured CoT prompting significantly improves accuracy primarily for high-resource Latin-script languages; and (3) controlled perturbations via negation and distractor sentences reduce model accuracy and attribution coherence. These findings highlight the limitations of CoT prompting, particularly in terms of multilingual robustness and interpretive transparency.

cs.CL↗

arXiv2025

Self-Ablating Transformers: More Interpretability, Less Sparsity

A growing intuition in machine learning suggests a link between sparsity and interpretability. We introduce a novel self-ablation mechanism to investigate this connection ante-hoc in the context of language transformers. Our approach dynamically enforces a k-winner-takes-all constraint, forcing the model to demonstrate selective activation across neuron and attention units. Unlike post-hoc methods that analyze already-trained models, our approach integrates interpretability directly into model training, promoting feature localization from inception. Training small models on the TinyStories dataset and employing interpretability tests, we find that self-ablation leads to more localized circuits, concentrated feature representations, and increased neuron specialization without compromising language modelling performance. Surprisingly, our method also decreased overall sparsity, indicating that self-ablation promotes specialization rather than widespread inactivity. This reveals a complex interplay between sparsity and interpretability, where decreased global sparsity can coexist with increased local specialization, leading to enhanced interpretability. To facilitate reproducibility, we make our code available at https://github.com/keenanpepper/self-ablating-transformers.

cs.LG↗

arXiv2025

World Model Agents with Change-Based Intrinsic Motivation

Sparse reward environments pose a significant challenge for reinforcement learning due to the scarcity of feedback. Intrinsic motivation and transfer learning have emerged as promising strategies to address this issue. Change Based Exploration Transfer (CBET), a technique that combines these two approaches for model-free algorithms, has shown potential in addressing sparse feedback but its effectiveness with modern algorithms remains understudied. This paper provides an adaptation of CBET for world model algorithms like DreamerV3 and compares the performance of DreamerV3 and IMPALA agents, both with and without CBET, in the sparse reward environments of Crafter and Minigrid. Our tabula rasa results highlight the possibility of CBET improving DreamerV3's returns in Crafter but the algorithm attains a suboptimal policy in Minigrid with CBET further reducing returns. In the same vein, our transfer learning experiments show that pre-training DreamerV3 with intrinsic rewards does not immediately lead to a policy that maximizes extrinsic rewards in Minigrid. Overall, our results suggest that CBET provides a positive impact on DreamerV3 in more complex environments like Crafter but may be detrimental in environments like Minigrid. In the latter case, the behaviours promoted by CBET in DreamerV3 may not align with the task objectives of the environment, leading to reduced returns and suboptimal policies.

cs.LG↗