pith. sign in

arxiv: 2405.08448 · v1 · pith:LC5NIJSYnew · submitted 2024-05-14 · 💻 cs.LG · cs.AI

Understanding the performance gap between online and offline alignment algorithms

classification 💻 cs.LG cs.AI
keywords offlinealgorithmsalignmentperformanceonlinesamplingclassificationdata
0
0 comments X
read the original abstract

Reinforcement learning from human feedback (RLHF) is the canonical framework for large language model alignment. However, rising popularity in offline alignment algorithms challenge the need for on-policy sampling in RLHF. Within the context of reward over-optimization, we start with an opening set of experiments that demonstrate the clear advantage of online methods over offline methods. This prompts us to investigate the causes to the performance discrepancy through a series of carefully designed experimental ablations. We show empirically that hypotheses such as offline data coverage and data quality by itself cannot convincingly explain the performance difference. We also find that while offline algorithms train policy to become good at pairwise classification, it is worse at generations; in the meantime the policies trained by online algorithms are good at generations while worse at pairwise classification. This hints at a unique interplay between discriminative and generative capabilities, which is greatly impacted by the sampling process. Lastly, we observe that the performance discrepancy persists for both contrastive and non-contrastive loss functions, and appears not to be addressed by simply scaling up policy networks. Taken together, our study sheds light on the pivotal role of on-policy sampling in AI alignment, and hints at certain fundamental challenges of offline alignment algorithms.

This paper has not been read by Pith yet.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 9 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Learn from Your Mistakes: Tree-like Self-Play for Secure Code LLMs

    cs.CR 2026-06 unverdicted novelty 6.0

    TSP reframes secure code generation as a tree-structured self-play process that supplies dense on-policy signals at vulnerability-prone nodes, yielding higher security pass rates and cross-language generalization than...

  2. Response Time Enhances Alignment with Heterogeneous Preferences

    cs.LG 2026-05 unverdicted novelty 6.0

    Response times modeled as drift-diffusion processes enable consistent estimation of population-average preferences from heterogeneous anonymous binary choices.

  3. When Errors Can Be Beneficial: A Categorization of Imperfect Rewards for Policy Gradient

    cs.LG 2026-04 unverdicted novelty 6.0

    Certain errors in proxy rewards for policy gradient methods can be benign or beneficial by preventing policies from stalling on outputs with mediocre ground truth rewards, enabling improved RLHF metrics and reward des...

  4. Bootstrapping Post-training Signals for Open-ended Tasks via Rubric-based Self-play on Pre-training Text

    cs.CL 2026-04 unverdicted novelty 6.0

    POP bootstraps post-training signals for open-ended LLM tasks by synthesizing rubrics during self-play on pretraining corpus, yielding performance gains on Qwen-2.5-7B across healthcare QA, creative writing, and instr...

  5. SAM 3D: 3Dfy Anything in Images

    cs.CV 2025-11 unverdicted novelty 6.0

    SAM 3D reconstructs 3D objects from single images with geometry, texture, and pose using human-model annotated data at scale and synthetic-to-real training, achieving 5:1 human preference wins.

  6. Beyond the 80/20 Rule: High-Entropy Minority Tokens Drive Effective Reinforcement Learning for LLM Reasoning

    cs.CL 2025-06 conditional novelty 6.0

    High-entropy minority tokens drive RLVR gains, so restricting gradients to the top 20% maintains or improves performance over full updates on Qwen3 models, especially larger ones.

  7. RLFTSim: Realistic and Controllable Multi-Agent Traffic Simulation via Reinforcement Learning Fine-Tuning

    cs.RO 2026-05 unverdicted novelty 5.0

    RLFTSim uses RL fine-tuning on a pre-trained model with a balanced reward to align traffic simulator rollouts to real data distributions and distill goal-conditioned controllability, reporting SOTA realism on the Waym...

  8. Advancing Multi-Agent RAG Systems with Minimalist Reinforcement Learning

    cs.CL 2025-05 unverdicted novelty 5.0

    Mujica-MyGo decomposes multi-turn RAG interactions via multi-agent workflows and applies minimalist policy gradient optimization to improve performance on QA benchmarks while avoiding long-context problems.

  9. Sharpness-Guided Group Relative Policy Optimization via Probability Shaping

    cs.LG 2025-10 unverdicted novelty 4.0

    GRPO-SG is a sharpness-guided token-weighted variant of GRPO that downweights high-gradient tokens to stabilize optimization and improve generalization in reinforcement learning with verifiable rewards.