REVIEW 9 cited by
WARP: On the Benefits of Weight Averaged Rewarded Policies
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
WARP: On the Benefits of Weight Averaged Rewarded Policies
read the original abstract
Reinforcement learning from human feedback (RLHF) aligns large language models (LLMs) by encouraging their generations to have high rewards, using a reward model trained on human preferences. To prevent the forgetting of pre-trained knowledge, RLHF usually incorporates a KL regularization; this forces the policy to remain close to its supervised fine-tuned initialization, though it hinders the reward optimization. To tackle the trade-off between KL and reward, in this paper we introduce a novel alignment strategy named Weight Averaged Rewarded Policies (WARP). WARP merges policies in the weight space at three distinct stages. First, it uses the exponential moving average of the policy as a dynamic anchor in the KL regularization. Second, it applies spherical interpolation to merge independently fine-tuned policies into a new enhanced one. Third, it linearly interpolates between this merged model and the initialization, to recover features from pre-training. This procedure is then applied iteratively, with each iteration's final model used as an advanced initialization for the next, progressively refining the KL-reward Pareto front, achieving superior rewards at fixed KL. Experiments with GEMMA policies validate that WARP improves their quality and alignment, outperforming other open-source LLMs.
Forward citations
Cited by 9 Pith papers
-
Extrapolative Weight Averaging Reveals Correctness-Efficiency Frontiers in Code RL
Extrapolative weight averaging of RL checkpoints trained under nested unit-test coverage extends a correctness-efficiency frontier and boosts ensemble pass rates in code generation across model scales and inference modes.
-
REVA-PO: Stabilizing Reinforcement Learning for Chest X-ray Report Generation
REVA-PO stabilizes GRPO-style RL for CXR report generation via response-level adaptive KL weights and validation-anchored policy resets, reporting new SOTA BLEU and clinical F1 scores.
-
Multi-Turn On-Policy Distillation with Prefix Replay
ReOPD offline-distills multi-turn agentic LLMs via teacher-prefix replay plus step-decay sampling, matching online OPD accuracy at ≥4× speed with zero tool calls.
-
EMAgnet: Parameter-Space EMA Regularization for Policy Gradient Self-Play in Large Games
EMAgnet replaces uniform-magnet regularization in PPO self-play with an EMA of last-iterate policy parameters and reports lower exploitability on most tested zero-sum benchmarks, especially those with dominated strategies.
-
Spectral Souping: A Unified Framework for Online Preference Alignment
Spectral Souping learns offline specialized policies for fine-grained preferences and merges them online using a discovered universal spectral representation for efficient LLM alignment.
-
Decouple before Integration: Test-time Synthesis of SFT and RLVR Task Vectors
DoTS decouples SFT and RLVR training then synthesizes their task vectors at inference time to match integrated training results at ~3% compute cost.
-
Robust Reward Modeling for Large Language Models via Causal Decomposition
A decoder-based reconstruction error regularizer improves reward model accuracy on RewardBench from 0.832 to 0.868 while favoring shorter, less sycophantic outputs.
-
CoLD: Counterfactually-Guided Length Debiasing for Process Reward Models in Mathematical Reasoning
CoLD mitigates length bias in process reward models for mathematical reasoning via counterfactual guidance, length penalties, bias estimation, and joint training, improving step selection accuracy and conciseness on M...
-
Model Merging in LLMs, MLLMs, and Beyond: Methods, Theories, Applications and Opportunities
The paper introduces a new taxonomy for model merging methods and reviews their applications in LLMs, MLLMs, continual learning, multi-task learning, and other subfields while outlining open challenges.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.