Collab: Controlled Decoding using Mixture of Agents for LLM Alignment

Authors: Souradip Chakraborty, Sujay Bhatt, Udari Sehwag, Soumya Suvra Ghosal, Jiahao Qiu, Mengdi Wang, Dinesh Manocha, Furong Huang, Alec Koppel, Sumitra Ganesh

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we present a comprehensive empirical analysis of our proposed framework, tested across various open-source datasets and state-of-the-art models (Lambert et al., 2024). Our findings demonstrate Collab s effectiveness in aligning language model outputs with specific target rewards. For implementation, we set the number of tokens sampled (top-p) p = 10 and the decoding alignment parameter α = 1. Reproducibility is ensured through the use of publicly available resources.
Researcher Affiliation Collaboration 1JPMorgan AI Research 2University of Maryland, College Park 3Princeton University
Pseudocode Yes Algorithm 1 Mixture of Agents based Controlled Decoding for LLM Alignment
Open Source Code No Reproducibility is ensured through the use of publicly available resources. This statement refers to resources used for experiments, not the authors' own code.
Open Datasets Yes 1. Evaluation-1 to Evaluation-4 (Task-I): For this task, we utilize the Berkeley Nectar dataset (Zhu et al., 2023) to test the agent s capacity for multi-turn dialogues and question answering. 2. Evaluation-5 to Evaluation-7 (Task-II): We employ the HH-RLHF dataset (Bai et al., 2022) to assess the agent s helpfulness and ethical alignment in response generation.
Dataset Splits No For evaluation, we compare the performance of the response generated by the language model corresponding to each prompt in the test dataset. Following (Khanov et al., 2024; Chakraborty et al., 2024b), we limit the maximum length of the prompt and generated continuation to 128 and 2048 tokens, respectively. The paper mentions using a "test dataset" but does not specify the explicit splits (e.g., percentages, counts) for the datasets used.
Hardware Specification Yes We run all experiments with Python 3.7.4 and Py Torch 1.9.0. For all experimentation, we use two Nvidia RTX A6000 GPUs.
Software Dependencies Yes We run all experiments with Python 3.7.4 and Py Torch 1.9.0.
Experiment Setup Yes For implementation, we set the number of tokens sampled (top-p) p = 10 and the decoding alignment parameter α = 1.