Detecting Proxy Gaming in RL and LLM Alignment via Evaluator Stress Tests (2026.findings-acl)
Copied to clipboard
| Challenge: | Proxy optimization is a challenge spanning reinforcement learning and LLM alignment. |
| Approach: | They propose an invariance-based framework that detects proxy gaming by separating exploitable sensitivity from content-driven improvements using semantic validity audits. |
| Outcome: | The proposed framework achieves 78.4% precision and 81.7% recall across 15 environments and 5 algorithms. |
Similar Papers
Rethinking the Role of Proxy Rewards in Language Model Alignment (2024.emnlp-main)
Copied to clipboard
| Challenge: | Typically, the human feedback is used to train a proxy reward model (RM), and a policy model is optimized over the reward signal from the RM using RL. |
| Approach: | They aim to replicate the ground truth (gold) reward signal by achieving a monotonic relationship between the proxy and gold reward signals after training the model using the proxy reward in reinforcement learning (RL). |
| Outcome: | The proposed model shows competitive performances with strong open-source RMs in alignment benchmarks. |
Semantic-Space Exploration and Exploitation in RLVR for LLM Reasoning (2026.findings-acl)
Copied to clipboard
Fanding Huang, Guanbo Huang, Xiao Fan, Yi He, Xiao Liang, Xiao Chen, Qinting Jiang, Faisal Nadeem Khan, Jingyan Jiang, Zhi Wang
| Challenge: | Recent advances in Reinforcement Learning with Verifiable Rewards (RLVR) have substantially improved the reasoning abilities of Large Language Models (LLMs). |
| Approach: | They propose a method that balances exploration and exploitation in the hidden-state space of response trajectories. |
| Outcome: | The proposed model yields consistent improvements across models, algorithms and reasoning benchmarks. |
Toward Scalable Verifiable Reward: Proxy State-Based Evaluation for Multi-turn Tool-Calling LLM Agents (2026.acl-industry)
Copied to clipboard
Yun-Shiuan Chuang, Chaitanya Kulkarni, Alec M. Chiu, Avinash Thangali, Zijie Pan, Shivani Shekhar, Yirou Ge, Yixi Li, Uma Kona, Linsey Pang, Prakhar Mehrotra
| Challenge: | Existing agentic benchmarks rely on deterministic backends and are costly to build and iterate. |
| Approach: | They propose a framework that preserves final state-based evaluation without a deterministic database. |
| Outcome: | The proposed framework produces stable, model-differentiating rankings across families and inference-time reasoning efforts. |
Model Consistency as a Cheap yet Predictive Proxy for LLM Elo Scores (2025.emnlp-main)
Copied to clipboard
| Challenge: | a rapid proliferation of large language models (LLMs) makes it difficult to assess which models are best suited for specific tasks. |
| Approach: | They find that the consistency of an LLM's Elo score is 91% correlated with its own human-produced Elo scores. |
| Outcome: | a new method to evaluate large language models is needed to scale with the increasing number of models released . current best way is to measure model's Elo score by comparing it to other models in contests . a simple proxy for Elo scores can be computed cheaply without human data or prior knowledge . |
Outcome Accuracy is Not Enough: Aligning the Reasoning Process of Reward Models (2026.acl-long)
Copied to clipboard
Binghai Wang, Yantao Liu, Yuxuan Liu, Tianyi Tang, Shenzhi Wang, Chang Gao, Chujie Zheng, Yichang Zhang, Le Yu, Shixuan Liu, Tao Gui, Qi Zhang, Xuanjing Huang, Bowen Yu, Fei Huang, Junyang Lin
| Challenge: | Recent studies observe a phenomenon where reward models achieve high accuracy on static datasets but fail to generalize effectively during RLHF. |
| Approach: | They propose a method that combines rationale consistency with outcome accuracy to improve performance on RM-Bench and JudgeBench. |
| Outcome: | The proposed method surpasses baselines on RM-Bench and JudgeBench by an average of 5% and improves creative writing tasks by 7%. |
Adversarial Preference Learning for Robust LLM Alignment (2025.findings-acl)
Copied to clipboard
Yuanfu Wang, Pengyu Wang, Chenyang Xi, Bo Tang, Junyi Zhu, Wenqiang Wei, Chen Chen, Chao Yang, Jingfeng Zhang, Chaochao Lu, Yijun Niu, Keming Mao, Zhiyu Li, Feiyu Xiong, Jie Hu, Mingchuan Yang
| Challenge: | Modern language models rely on Reinforcement Learning from Human Feedback (RLHF) to encourage safe behaviors, but they remain vulnerable to adversarial attacks due to three key limitations: (1) the inefficiency and high cost of human annotation; (2) the vast diversity of potential adversarials; and (3) the risk of feedback bias and reward hacking. |
| Approach: | They propose an iterative adversarial training method that incorporates three key innovations to address these challenges. |
| Outcome: | Experiments on Mistral-7B-Instruct-v0.3 show that the proposed method significantly enhances robustness and reduces harmful outputs from 5.88% to 0.43%. |
Rewarding Semantic Similarity under Optimized Alignments for AMR-to-Text Generation (2022.acl-short)
Copied to clipboard
| Challenge: | Automatic evaluation metrics score natural language generation systems based on how well they lexically align to humanannotated references. |
| Approach: | They propose to replace greedy alignments in BERTScore with optimized ones that replace the n-gram matching BERTAcore metrics with a token embedding to prevent domain mismatch. |
| Outcome: | The proposed metrics outperform cross-entropy and BLEU reward baselines on AMR-to-text generation. |
MASPO: Unifying Gradient Utilization, Probability Mass, and Signal Reliability for Robust and Sample-Efficient LLM Reasoning (2026.acl-long)
Copied to clipboard
Xiaoliang Fu, Jiaye Lin, Yangyi Fang, Binbin Zheng, Chaowen Hu, Zekai Shao, Cong Qin, Lu Pan, Ke Zeng, Xunliang Cai
| Challenge: | Existing RLVR algorithms rely on rigid, uniform, and symmetric trust region mechanisms . current algorithms lack robustness, asymmetric signal reliability and inefficient gradient utilization . |
| Approach: | They propose a framework to harmonize three dimensions of RLVR algorithms, a paper argues . a binary cutoff is used to discard valuable reinforcement signals, they argue . |
| Outcome: | The proposed framework outperforms baselines in evaluating a robust RLVR solution. |
TriPlay-RL: Tri-Role Self-Play Reinforcement Learning for LLM Safety Alignment (2026.acl-long)
Copied to clipboard
Zhewen Tan, Wenhan Yu, Jianfeng Si, Tongxin Liu, Kaiqi Guan, Huiyan Jin, Jiawen Tao, Xiaokun Yuan, Xiangzheng Zhang, Duohe Ma, Tong Yang, Lin Sun
| Challenge: | Existing approaches to safety alignment of large language models rely on costly manual annotations or human review. |
| Approach: | They propose a closed-loop reinforcement learning framework called TriPlay-RL that enables iterative collaboration among three roles with near-zero manual annotation. |
| Outcome: | The proposed framework achieves 20%–50% improvement in adversarial effectiveness while preserving high output diversity while achieving 10%–30% gains in safety performance without degrading general reasoning capability. |
Adversarial Preference Optimization: Enhancing Your Alignment via RM-LLM Game (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for training large language models require additional annotations to adjust to shifted distributions. |
| Approach: | They propose an algorithm that allows LLMs and reward models to update alternatively via a min-max game to improve their alignment. |
| Outcome: | The proposed framework improves existing alignment baselines in terms of LLM helpfulness and harmlessness. |