arXiv subjects

Aakash Sen Sharma

Publications and source records attributed to Aakash Sen Sharma.

4 recordsLinked to original sources

arXiv2025

The Realignment Problem: When Right becomes Wrong in LLMs

Post-training alignment of large language models (LLMs) relies on large-scale human annotations guided by policy specifications that change over time. Cultural shifts, value reinterpretations, and regulatory or industrial updates make static alignment increasingly brittle. As policies evolve, deployed models can diverge from current alignment objectives, creating an Alignment-Reality Gap that is difficult to audit or correct. Existing remediation typically requires re-annotation under revised guidelines, which introduces systematic challenges, including guideline ambiguity, annotator interpretation drift, and reduced consistency at scale. We introduce TRACE (Triage and Re-align by Alignment Conflict Evaluation), a framework that transforms realignment into a structured optimization problem over existing data without requiring fresh human annotation. Leveraging a stronger model as a proxy judge, TRACE operates via a three-stage pipeline: (1) triaging preference pairs into inversion, suppression, or retention categories based on alignment conflicts; (2) computing an alignment impact score via bi-level optimization to prioritize high-leverage samples; and (3) executing updates using a hybrid objective that combines relational losses (e.g., IPO) for preference inversion and punitive losses (e.g., NPO) for response suppression. Experiments on Qwen2.5-7B, Gemma-2-9B, and Llama-3.1-8B demonstrate robust realignment on synthetic benchmarks and the PKU-SafeRLHF dataset without degrading general utility. This work provides a scalable approach for LLM realignment under evolving data annotation policies and alignment guidelines. We release our code: https://respailab.github.io/TRACE/

cs.CL↗

arXiv2025

Policy Optimization Prefers The Path of Least Resistance

Policy optimization (PO) algorithms are used to refine Large Language Models for complex, multi-step reasoning. Current state-of-the-art pipelines enforce a strict think-then-answer format to elicit chain-of-thought (CoT); however, the behavior of PO when these rigid constraints are relaxed into an open-ended CoT structure remains an under-studied question. We investigate this gap with an extensive suite of controlled experiments and identify a consistent principle: \textit{policy optimization consistently follows the path of least resistance}. When afforded the flexibility to interleave reasoning and response, policy optimization consistently learns to discard explicit reasoning, causing the policy to degenerate to a direct \texttt{ }-only format. This outcome holds true across various models and algorithms. We find that this collapse in format is persistent even when the complex \texttt{ } format is assigned up to 4x larger reward weights. We formalize this principle through a series of controlled reward decomposition experiments, demonstrating a clear hierarchy: PO systematically optimizes for the simplest reward component first, a preference that holds even when faced with mutually exclusive choices or strong incentives for more complex behaviors. Finally, we show that successful convergence on the high-reward shortcut is not a low-effort drift but is driven by the optimization process that requires the KL-regularized policy to have sufficient freedom to make a significant shift from its initial prior. Our findings reveal that granting policies the freedom to diverge is a double-edged sword: while necessary for discovering high-reward shortcuts, it also creates a powerful incentive to game the simplest aspects of the reward function, posing a critical challenge for reward hacking under alignment.

cs.CL↗

arXiv2025

Nine Ways to Break Copyright Law and Why Our LLM Won't: A Fair Use Aligned Generation Framework

Large language models (LLMs) commonly risk copyright infringement by reproducing protected content verbatim or with insufficient transformative modifications, posing significant ethical, legal, and practical concerns. Current inference-time safeguards predominantly rely on restrictive refusal-based filters, often compromising the practical utility of these models. To address this, we collaborated closely with intellectual property experts to develop FUA-LLM (Fair Use Aligned Language Models), a legally-grounded framework explicitly designed to align LLM outputs with fair-use doctrine. Central to our method is FairUseDB, a carefully constructed dataset containing 18,000 expert-validated examples covering nine realistic infringement scenarios. Leveraging this dataset, we apply Direct Preference Optimization (DPO) to fine-tune open-source LLMs, encouraging them to produce legally compliant and practically useful alternatives rather than resorting to blunt refusal. Recognizing the shortcomings of traditional evaluation metrics, we propose new measures: Weighted Penalty Utility and Compliance Aware Harmonic Mean (CAH) to balance infringement risk against response utility. Extensive quantitative experiments coupled with expert evaluations confirm that FUA-LLM substantially reduces problematic outputs (up to 20\%) compared to state-of-the-art approaches, while preserving real-world usability.

cs.CL↗

arXiv2024

Unlearning or Concealment? A Critical Analysis and Evaluation Metrics for Unlearning in Diffusion Models

Recent research has seen significant interest in methods for concept removal and targeted forgetting in text-to-image diffusion models. In this paper, we conduct a comprehensive white-box analysis showing the vulnerabilities in existing diffusion model unlearning methods. We show that existing unlearning methods lead to decoupling of the targeted concepts (meant to be forgotten) for the corresponding prompts. This is concealment and not actual forgetting, which was the original goal. This paper presents a rigorous theoretical and empirical examination of five commonly used techniques for unlearning in diffusion models, while showing their potential weaknesses. We introduce two new evaluation metrics: Concept Retrieval Score (\textbf{CRS}) and Concept Confidence Score (\textbf{CCS}). These metrics are based on a successful adversarial attack setup that can recover \textit{forgotten} concepts from unlearned diffusion models. \textbf{CRS} measures the similarity between the latent representations of the unlearned and fully trained models after unlearning. It reports the extent of retrieval of the \textit{forgotten} concepts with increasing amount of guidance. CCS quantifies the confidence of the model in assigning the target concept to the manipulated data. It reports the probability of the \textit{unlearned} model's generations to be aligned with the original domain knowledge with increasing amount of guidance. The \textbf{CCS} and \textbf{CRS} enable a more robust evaluation of concept erasure methods. Evaluating existing five state-of-the-art methods with our metrics, reveal significant shortcomings in their ability to truly \textit{unlearn}. Source Code: \color{blue}{https://respailab.github.io/unlearning-or-concealment}

cs.LG↗