arXiv subjects

Ujwal Dinesha

Publications and source records attributed to Ujwal Dinesha.

4 recordsLinked to original sources

arXiv2025

PITA: Preference-Guided Inference-Time Alignment for LLM Post-Training

Inference-time alignment enables large language models (LLMs) to generate outputs aligned with end-user preferences without further training. Recent post-training methods achieve this by using small guidance models to modify token generation during inference. These methods typically optimize a reward function KL-regularized by the original LLM taken as the reference policy. A critical limitation, however, is their dependence on a pre-trained reward model, which requires fitting to human preference feedback--a potentially unstable process. In contrast, we introduce PITA, a novel framework that integrates preference feedback directly into the LLM's token generation, eliminating the need for a reward model. PITA learns a small preference-based guidance policy to modify token probabilities at inference time without LLM fine-tuning, reducing computational cost and bypassing the pre-trained reward model dependency. The problem is framed as identifying an underlying preference distribution, solved through stochastic search and iterative refinement of the preference-based guidance model. We evaluate PITA across diverse tasks, including mathematical reasoning and sentiment classification, demonstrating its effectiveness in aligning LLM outputs with user preferences.

cs.AI↗

arXiv2025

DOPL: Direct Online Preference Learning for Restless Bandits with Preference Feedback

Restless multi-armed bandits (RMAB) has been widely used to model constrained sequential decision making problems, where the state of each restless arm evolves according to a Markov chain and each state transition generates a scalar reward. However, the success of RMAB crucially relies on the availability and quality of reward signals. Unfortunately, specifying an exact reward function in practice can be challenging and even infeasible. In this paper, we introduce Pref-RMAB, a new RMAB model in the presence of \textit{preference} signals, where the decision maker only observes pairwise preference feedback rather than scalar reward from the activated arms at each decision epoch. Preference feedback, however, arguably contains less information than the scalar reward, which makes Pref-RMAB seemingly more difficult. To address this challenge, we present a direct online preference learning (DOPL) algorithm for Pref-RMAB to efficiently explore the unknown environments, adaptively collect preference data in an online manner, and directly leverage the preference feedback for decision-makings. We prove that DOPL yields a sublinear regret. To our best knowledge, this is the first algorithm to ensure $\tilde{\mathcal{O}}(\sqrt{T\ln T})$ regret for RMAB with preference feedback. Experimental results further demonstrate the effectiveness of DOPL.

cs.LG↗

arXiv2025

Risk-Averse Finetuning of Large Language Models

We consider the challenge of mitigating the generation of negative or toxic content by the Large Language Models (LLMs) in response to certain prompts. We propose integrating risk-averse principles into LLM fine-tuning to minimize the occurrence of harmful outputs, particularly rare but significant events. By optimizing the risk measure of Conditional Value at Risk (CVaR), our methodology trains LLMs to exhibit superior performance in avoiding toxic outputs while maintaining effectiveness in generative tasks. Empirical evaluations on sentiment modification and toxicity mitigation tasks demonstrate the efficacy of risk-averse reinforcement learning with human feedback (RLHF) in promoting a safer and more constructive online discourse environment.

cs.AI↗

arXiv2023

EdgeRIC: Empowering Realtime Intelligent Optimization and Control in NextG Networks

Radio Access Networks (RAN) are increasingly softwarized and accessible via data-collection and control interfaces. RAN intelligent control (RIC) is an approach to manage these interfaces at different timescales. In this paper, we develop a RIC platform called RICworld, consisting of (i) EdgeRIC, which is colocated, but decoupled from the RAN stack, and can access RAN and application-level information to execute AI-optimized and other policies in realtime (sub-millisecond) and (ii) DigitalTwin, a full-stack, trace-driven emulator for training AI-based policies offline. We demonstrate that realtime EdgeRIC operates as if embedded within the RAN stack and significantly outperforms a cloud-based near-realtime RIC (> 15 ms latency) in terms of attained throughput. We train AI-based polices on DigitalTwin, execute them on EdgeRIC, and show that these policies are robust to channel dynamics, and outperform queueing-model based policies by 5% to 25% on throughput and application-level benchmarks in a variety of mobile environments.

cs.NI↗