EvalStop: Using World Feedback to Detect and Correct Reward Overoptimization in Multi-Tenant RLHF Platforms
Authors: Guilin Zhang, Chuanyi Sun, Kai Zhao, Xu Chu, Shahryar Sarkani, John M. Fossaceca
Abstract
Cloud LLM fine-tuning platforms increasingly serve RLHF workloads, where a learned reward model is optimized as a proxy for human quality. As Gao et al. (2023) showed, this proxy diverges from world feedback (downstream eval metrics) under sustained optimization pressure, a phenomenon known as reward overoptimization. Existing platform schedulers ignore this divergence: non-clairvoyant schedulers optimize JCT without any quality signal, SLAQ-style quality-aware schedulers use training loss (a weaker proxy that drops monotonically through hacking), and classical per-job early stopping requires human monitoring and does not free shared GPUs. We propose EvalStop, a composable scheduling primitive that terminates jobs on k consecutive eval-score declines, releases GPUs, preserves the best checkpoint, and delegates to any base scheduler. We frame scheduler-level early stopping as a detection problem and evaluate it in a discrete-event simulator whose RLHF workload mixes reward-hacking and structurally healthy runs, with ground-truth labels hidden from schedulers. On RLHF-heavy workloads (80% RLHF, 64 GPUs), EvalStop achieves precision 98% / recall 99% / FPR 1.5% while improving JCT by 9% and cutting wasted compute by 22% over SRTF-Est (p<0.05). Trivial fixed-progress and loss-plateau competitors either incur 65% FPR on healthy RLHF or miss over half of true hacking cases. Gains compose across every base scheduler tested (9-25% JCT) and detection quality stays stable under eval noise (precision at least 91% at noise std <= 0.05) and hacking base rate (precision at least 89% across 20-80% hacking fractions).
RLHF evaluation should track how failures emerge, where they localize, and which warning signals appear before external quality degrades. We study this problem with a compact RLHF pipeline built for this paper, including PPO, DPO, uncertainty-penalized PPO (UP-PPO), reward-model uncertainty, approximate policy drift, diversity and repetition diagnostics, and two external LLM judges. Rather than treating reward hacking as a single terminal event, we classify matched checkpoint and prompt-level transitions by the directions of learned reward R_phi, judge scores R_dag and R2_dag, and their average R_dag. The main empirical findings are that aggressive PPO produces the clearest localized reward-hacking signal, UP-PPO reduces but does not eliminate that signal, row-level diagnostics reveal failures hidden by checkpoint averages, and pre-transition features partially anticipate future localized reward hacking. The central conclusion is methodological: RLHF failures are training dynamics that can be classified, localized, and partially anticipated, not only final-model pathologies. The repository is available at github.com/zabahana/rlhf-failure-modes-diagnostics. The pipeline is also deployed as a live interactive web demo for model comparison and diagnostic views at rlhf-failures.zelalem.ai.
Zelalem Abahana, David Evans, Satish Mahadevan Srinivasan +1
Supervised fine-tuning (SFT) can equip large language models (LLMs) with domain knowledge for high-performance computing (HPC) tasks such as data race detection and benchmark question answering. However, knowledge alone does not guarantee task-appropriate behavior: the same SFT model that correctly classifies 88.65% of C/C++ data race samples produces verbose, imprecise answers to factual queries, with 65.9% of MLPerf responses exceeding 40 characters. Reinforcement learning (RL) post-training addresses this gap by optimizing for task-specific rewards rather than token-level imitation. Yet HPC tasks exhibit extreme heterogeneity, with binary classification, factual QA, and semantic generation differing by 58x in answer length, spanning three distinct reward distributions, and showing widely varying SFT accuracy. This makes uniform-weight RL methods such as GRPO suboptimal. We propose HARGO, Heterogeneity-Aware Reward-Guided Optimization, which introduces per-response importance weighting via confidence-modulated advantage: computing a discrimination signal from group-level reward contrast and a confidence signal from reference model log-probabilities, then modulating the advantage before computing per-response weights, without requiring task-type labels. Across four HPC tasks and nine methods, HARGO achieves the best performance on all three primary metrics: WinRate 54.62%, Data Race F1 91.30%, and PLP Similarity 0.8558. Ablation confirms complementary contributions from both signals. HARGO establishes the best overall alignment quality among compared methods for heterogeneous HPC tasks.
Reinforcement learning with verifiable rewards (RLVR) is bottlenecked by rollout generation, yet many sampled prompts produce saturated groups (all responses correct or all incorrect) whose zero reward variance yields no policy-gradient signal. Existing remedies either oversample a larger candidate pool and discard saturated prompts (dynamic sampling), paying heavy extra rollouts, or predict prompt difficulty before sampling, which is fragile under a shifting policy. We observe that a group's effectiveness is usually decided early, within the first few of its rollouts, so spending a full group on an already-decided prompt is wasteful. We cast per-step rollout collection as a budget-constrained sequential allocation (optimal stopping) problem and introduce SARA (Sequential Adaptive Rollout Allocation). SARA maintains a Beta posterior over each prompt's success rate, evaluates a closed-form predictor of group effectiveness, and applies a two-threshold, SPRT-style rule that commits effective groups, abandons saturated ones after a short probe, and reallocates the freed budget to fresh prompts, without any extra prediction rollouts. We prove abandonment reliability, expected rollout savings, fixed-budget yield dominance, and a link between effective-group yield and the GRPO gradient norm. On mathematical reasoning and planning with 1.5B/3B models on a single GPU, SARA matches DPS (both below the DS oracle) while using 22% fewer rollouts than DS; composing SARA with DPS yields the best accuracy, slightly above DS, at 67% fewer rollouts (near-uniform cost).