REVIEW 12 cited by
Interpretable Preferences via Multi-Objective Reward Modeling and Mixture-of-Experts
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
Interpretable Preferences via Multi-Objective Reward Modeling and Mixture-of-Experts
read the original abstract
Reinforcement learning from human feedback (RLHF) has emerged as the primary method for aligning large language models (LLMs) with human preferences. The RLHF process typically starts by training a reward model (RM) using human preference data. Conventional RMs are trained on pairwise responses to the same user request, with relative ratings indicating which response humans prefer. The trained RM serves as a proxy for human preferences. However, due to the black-box nature of RMs, their outputs lack interpretability, as humans cannot intuitively understand why an RM thinks a response is good or not. As RMs act as human preference proxies, we believe they should be human-interpretable to ensure that their internal decision processes are consistent with human preferences and to prevent reward hacking in LLM alignment. To build RMs with interpretable preferences, we propose a two-stage approach: i) train an Absolute-Rating Multi-Objective Reward Model (ArmoRM) with multi-dimensional absolute-rating data, each dimension corresponding to a human-interpretable objective (e.g., honesty, verbosity, safety); ii) employ a Mixture-of-Experts (MoE) strategy with a gating network that automatically selects the most suitable reward objectives based on the context. We efficiently trained an ArmoRM with Llama-3 8B and a gating network consisting of a shallow MLP on top of the ArmoRM. Our trained model, ArmoRM-Llama3-8B, obtains state-of-the-art performance on RewardBench, a benchmark evaluating RMs for language modeling. Notably, the performance of our model surpasses the LLM-as-a-judge method with GPT-4 judges by a margin, and approaches the performance of the much larger Nemotron-4 340B reward model.
Forward citations
Cited by 12 Pith papers
-
On the Overscaling Curse of Parallel Thinking: System Efficacy Contradicts Sample Efficiency
Parallel thinking in LLMs suffers from overscaling where fixed global budgets waste samples; LanBo predicts per-sample budgets from latent states to raise utilization without hurting accuracy.
-
Chain-of-Models: Cross-Model Auditing for Bias-Robust LLM Judges
Per-bias selection of a cross-family LLM auditor lifts biased-judgment accuracy from 0.805/0.824 baselines to 0.884.
-
Explaining and Breaking the Safety-Helpfulness Ceiling via Preference Dimensional Expansion
MORA breaks the safety-helpfulness ceiling in LLMs by pre-sampling single-reward prompts and rewriting them to incorporate multi-dimensional intents, delivering 5-12.4% gains in sequential alignment and 4.6% overall i...
-
Explaining and Breaking the Safety-Helpfulness Ceiling via Preference Dimensional Expansion
MORA breaks the safety-helpfulness trade-off in LLM alignment by pre-sampling single-reward prompts and rewriting them to expand multi-dimensional reward diversity, yielding 5-12.4% single-preference gains in sequenti...
-
When Errors Can Be Beneficial: A Categorization of Imperfect Rewards for Policy Gradient
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...
-
OpsLLM: Construction of Large Language Model for Software Operations with Multi-stage Learning
OpsLLM's pipeline (HITL data curation, SFT, GRPO RL with a domain process reward model) improves LLM accuracy on software-operations QA and RCA, especially on in-distribution root-cause-analysis tasks.
-
Simultaneous Multi-objective Alignment Across Verifiable and Non-verifiable Rewards
MAHALO aligns LLMs to multiple objectives in one model via per-objective action heads and PRM-guided decoding, improving math, value, and tutoring metrics jointly.
-
RewardBench 2: Advancing Reward Model Evaluation
RewardBench 2 is a new benchmark that supplies challenging fresh human prompts for reward model evaluation, yielding lower average scores but higher correlation with downstream best-of-N sampling and RLHF training per...
-
GEOALIGN: Geometric Rollout Curation for Robust LLM Reinforcement Learning
Geoalign curates rollouts in online LLM RL by learning a projector on hidden states to flag and replace directionally inconsistent examples, yielding higher final performance and less oscillation than baselines on dia...
-
OpsLLM: Construction of Large Language Model for Software Operations with Multi-stage Learning
OpsLLM outperforms general LLMs on software operations QA and RCA tasks through human-in-the-loop data curation, supervised fine-tuning, and domain-specific reinforcement learning.
-
OpsLLM: Construction of Large Language Model for Software Operations with Multi-stage Learning
OpsLLM is a domain-specific LLM for software ops QA and RCA built with human-curated data, SFT, and RL using a domain process reward model, showing accuracy gains of 0.2-5.7% on QA and 2.7-70.3% on RCA over general LLMs.
-
T-POP: Test-Time Personalization with Online Preference Feedback
T-POP uses dueling-bandit token selection to learn a reward function online from pairwise user feedback, enabling test-time personalization of a frozen LLM without fine-tuning.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.