Publications

NeuroFaith: Evaluating LLM Self-Explanation Faithfulness via Internal Representation Alignment
Milan Bhan
Jean-Noël Vittaut
Nicolas Chesneau
A. Chandar
Marie-Jeanne Lesot
Large Language Models (LLMs) can generate plausible free text self-explanations to justify their answers. However, these natural language ex… (voir plus)planations may not accurately reflect the model's actual reasoning process, pinpointing a lack of faithfulness. Existing faithfulness evaluation methods rely primarily on behavioral tests or computational block analysis without examining the semantic content of internal neural representations. This paper proposes NeuroFaith, a flexible framework that measures the faithfulness of LLM free text self-explanation by identifying key concepts within explanations and mechanistically testing whether these concepts actually influence the model's predictions. We show the versatility of NeuroFaith across 2-hop reasoning and classification tasks. Additionally, we develop a linear faithfulness probe based on NeuroFaith to detect unfaithful self-explanations from representation space and improve faithfulness through steering. NeuroFaith provides a principled approach to evaluating and enhancing the faithfulness of LLM free text self-explanations, addressing critical needs for trustworthy AI systems.
2025-06-09
ArXiv (prépublication)
arxiv.org
arxiv.org
Parity Requires Unified Input Dependence and Negative Eigenvalues in SSMs
Behnoush Khavari
Mehran Shakerinava
Jayesh Khullar
Jerry Huang
Franccois Rivest
Siamak Ravanbakhsh
A. Chandar
2025-06-09
ICML.cc/2025/Workshop/MOSS (publié)
doi.org
openreview.net
Preservice Teachers’ Computational Thinking Profiles
Tanya Chichekian
Maria Cutumisu
Annie Savard
2025-06-09
Proceedings (International Conference of the Learning Sciences) (publié)
doi.org
Robust Reward Modeling via Causal Rubrics
Pragya Srivastava
Harman Singh
Rahul Madhavan
Gandharv Patil
Sravanti Addepalli
Arun Suggala
Rengarajan Aravamudhan
Soumya Sharma
Anirban Laha
Aravindan Raghuveer
Karthikeyan Shanmugam
Doina Precup
Reward models (RMs) for LLM alignment often exhibit reward hacking, mistaking spurious correlates (e.g., length, format) for causal quality … (voir plus)drivers (e.g., factuality, relevance), leading to brittle RMs. We introduce CROME (Causally Robust Reward Modeling), a causally-grounded framework using targeted augmentations to mitigate this. CROME employs: (1) Causal Augmentations, pairs isolating specific causal attribute changes, to enforce sensitivity, and (2) Neutral Augmentations, tie-labeled pairs varying spurious attributes while preserving causal content, to enforce invariance. Crucially, augmentations target LLM-identified causal rubrics, requiring no prior knowledge of spurious factors. CROME significantly outperforms baselines on RewardBench (Avg +5.4\%, Safety +13.2\%, Reasoning +7.2\%) and demonstrates enhanced robustness via improved Best-of-N performance across RewardBench, WildGuardTest, and GSM8k.
2025-06-09
ICML.cc/2025/Workshop/MoFA (poster)
doi.org
openreview.net
A Systematic Literature Review of Large Language Model Applications in the Algebra Domain
Yajie Song
Yimei Zhang
Doina Precup
Reihaneh Rabbany
Maria Cutumisu
2025-06-09
Proceedings (International Conference of the Learning Sciences) (publié)
doi.org
Test Time Adaptation Using Adaptive Quantile Recalibration
Paria Mehrbod
Pedro Vianna
Geraldin Nanfack
Guy Wolf
Eugene Belilovsky
2025-06-09
ICML.cc/2025/Workshop/PUT (poster)
doi.org
openreview.net
What Matters when Modeling Human Behavior using Imitation Learning?
Aneri Muni
Esther Derman
Vincent Taboga
Pierre-Luc Bacon
Erick Delage
As AI systems become increasingly embedded in human decision-making process, aligning their behavior with human values is critical to ensuri… (voir plus)ng safe and trustworthy deployment. A central approach to AI Alignment called Imitation Learning (IL), trains a learner to directly mimic desirable human behaviors from expert demonstrations. However, standard IL methods assume that (1) experts act to optimize expected returns; (2) expert policies are Markovian. Both assumptions are inconsistent with empirical findings from behavioral economics, according to which humans are (1) risk-sensitive; and (2) make decisions based on past experience. In this work, we examine the implications of risk sensitivity for IL and show that standard approaches do not capture all optimal policies under risk-sensitive decision criteria. By characterizing these expert policies, we identify key limitations of existing IL algorithms in replicating expert performance in risk-sensitive settings. Our findings underscore the need for new IL frameworks that account for both risk-aware preferences and temporal dependencies to faithfully align AI behavior with human experts.
2025-06-09
ICML.cc/2025/Workshop/MoFA (poster)
openreview.net
openreview.net
Improving Context Fidelity via Native Retrieval-Augmented Reasoning
Suyuchen Wang
Jinlin Wang
Xinyu Wang
Shiqi Li
Xiangru Tang
Sirui Hong
Xiao-Wen Chang
Chenglin Wu
Bang Liu
Large language models (LLMs) often struggle with context fidelity, producing inconsistent answers when responding to questions based on prov… (voir plus)ided information. Existing approaches either rely on expensive supervised fine-tuning to generate evidence post-answer or train models to perform web searches without necessarily improving utilization of the given context. We propose CARE, a novel native retrieval-augmented reasoning framework that teaches LLMs to explicitly integrate in-context evidence within their reasoning process with the model's own retrieval capabilities. Our method requires minimal labeled evidence data while significantly enhancing both retrieval accuracy and answer generation performance through strategically retrieved in-context tokens in the reasoning chain. Extensive experiments on multiple real-world and counterfactual QA benchmarks demonstrate that our approach substantially outperforms supervised fine-tuning, traditional retrieval-augmented generation methods, and external retrieval solutions. This work represents a fundamental advancement in making LLMs more accurate, reliable, and efficient for knowledge-intensive tasks.
2025-06-08
ICML.cc/2025/Workshop/LCFM (accepté)
openreview.net
openreview.net
A Meta-Learning Approach to Causal Inference
Dragos Cristian Manta
Philippe Brouillard
Dhanya Sridhar
Predicting the effect of unseen interventions is at the heart of many scientific endeavours. While causal discovery is often used to answer … (voir plus)these causal questions, it involves learning a full causal model, not tailored to the specific goal of predicting unseen interventions, and operates under stringent assumptions. We introduce a novel method based on meta-learning that predicts interventional effects without explicitly assuming a causal model. Our preliminary results on synthetic data show that it can provide good generalization to unseen interventions, and it even compares favorably to a causal discovery method. Our model-agnostic method opens up many avenues for future exploration, particularly for settings where causal discovery cannot be applied.
2025-06-08
ICML.cc/2025/Workshop/SIM (poster)
openreview.net
openreview.net
PyLO: Towards Accessible Learned Optimizers in PyTorch
Paul Janson
Benjamin Therien
Quentin Gregory Anthony
Xiaolong Huang
Abhinav Moudgil
Eugene Belilovsky
Learned optimizers have been an active research topic over the past decade, with increasing progress toward practical, general-purpose optim… (voir plus)izers that can serve as drop-in replacements for widely used methods like Adam. However, recent advances -- such as VeLO, which was meta-trained for 4000 TPU-months -- remain largely inaccessible to the broader community, in part due to their reliance on JAX and the absence of user-friendly packages for applying the optimizers after meta-training. To address this gap, we introduce PyLO, a PyTorch-based library that brings learned optimizers to the broader machine learning community through familiar, widely adopted workflows. Unlike prior work focused on synthetic or convex tasks, our emphasis is on applying learned optimization to real-world large-scale pre-training tasks. Our release includes a CUDA-accelerated version of the small_fc_lopt learned optimizer architecture from (Metz et al., 2022a), delivering substantial speedups -- from 39.36 to 205.59 samples/sec throughput for training ViT B/16 with batch size 32. PyLO also allows us to easily combine learned optimizers with existing optimization tools such as learning rate schedules and weight decay. When doing so, we find that learned optimizers can substantially benefit. Our code is available at https://github.com/Belilovsky-Lab/pylo
2025-06-08
ICML.cc/2025/Workshop/CODEML (publié)
doi.org
openreview.net
Quantized Disentanglement: A Practical Approach
Vitória Barin-Pacela
Kartik Ahuja
Simon Lacoste-Julien
P Vincent
2025-06-08
ICML.cc/2025/Workshop/SIM (poster)
openreview.net
openreview.net
Revisiting the Goldilocks Zone in Inhomogeneous Networks
Zacharie Garnier Cuchet
A. Chandar
Ekaterina Lobacheva
We investigate how architectural inhomogeneities—such as biases, layer normalization, and residual connections—affect the curvature of t… (voir plus)he loss landscape at initialization and its link to trainability. We focus on the Goldilocks zone, a region in parameter space with excess positive curvature, previously associated with improved optimization in homogeneous networks. To extend this analysis, we compare two scaling strategies: weight scaling and softmax temperature scaling. Our results show that in networks with biases or residual connections, both strategies identify a Goldilocks zone aligned with better training. In contrast, layer normalization leads to lower or negative curvature, yet stable optimization—revealing a disconnect between curvature and trainability. Softmax temperature scaling behaves more consistently across models, making it a more robust probe. Overall, the Goldilocks zone remains relevant in inhomogeneous networks, but its geometry and predictive power depend on architectural choices, particularly normalization.
2025-06-08
ICML.cc/2025/Workshop/HiLD (poster)
openreview.net
openreview.net