Small Reward Models via Backward Inference
cs.CL
Submitted: 2026-02-14
Updated: 2026-08-31
Code: https://github.com/yikee/FLIP
Project page: https://allenai.github.io/open-instruct
License: http://creativecommons.org/licenses/by/4.0/
The gist: Reward models (RMs) play a central role throughout the language model (LM) pipeline, particularly in non-verifiable domains.
Terminology
Abstract
Reward models (RMs) play a central role throughout the language model (LM) pipeline, particularly in non-verifiable domains. However, the dominant LLM-as-a-Judge paradigm relies on the strong reasoning capabilities of large models, while alternative approaches require reference responses or explicit rubrics, limiting flexibility and broader accessibility. In this work, we propose FLIP (FLipped Inference for Prompt reconstruction), a reference-free and rubric-free reward modeling approach that reformulates reward modeling through backward inference: inferring the instruction that would most plausibly produce a given response. The similarity between the inferred and the original instructions is then used as the reward signal. Evaluations across four domains using 13 small language models show that FLIP outperforms LLM-as-a-Judge baselines by an average of 79.6%. Moreover, FLIP substantially improves downstream performance in extrinsic evaluations under test-time scaling via parallel sampling and GRPO training. We further find that FLIP is particularly effective for longer outputs and robust to common forms of reward hacking. By explicitly exploiting the validation-generation gap, FLIP enables reliable reward modeling in downscaled regimes where judgment methods fail. Code available at https://github.com/yikee/FLIP.
Sources
- Reverse Engineering Human Preferences with Reinforcement Learning
- Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback
- Small Language Models are the Future of Agentic AI
- Large Language Monkeys: Scaling Inference Compute with Repeated Sampling
- BLEUBERI: BLEU is a surprisingly effective reward for instruction following
- Humans or LLMs as the Judge? A Study on Judgement Biases
- LongForm: Effective Instruction Tuning with Reverse Instructions
- REInstruct: Building Instruction Data from Unlabeled Corpus
- Gen-Z: Generative Zero-Shot Text Classification with Contextualized Label Descriptions
- Tulu 3: Pushing Frontiers in Open Language Model Post-Training
- QA-LIGN: Aligning LLMs through Constitutionally Decomposed QA
- Preference Leakage: A Contamination Problem in LLM-as-a-judge
- Model Swarms: Collaborative Search to Adapt LLM Experts via Swarm Intelligence
- Reverse Prompt Engineering
- Position: Enough of Scaling LLMs! Lets Focus on Downscaling
- Self-Alignment with Instruction Backtranslation
- A Survey on LLM-as-a-Judge
- Benchmarking and Improving Generator-Validator Consistency of Language Models
- Rubrics as Rewards: Reinforcement Learning Beyond Verifiable Domains
- Is Best-of-N the Best of Them? Coverage, Scaling, and Optimality in Inference-Time Alignment
Related papers
- Exploring Solution Divergence and Its Effect on Large Language Model Problem Solving
- Ishigaki-IDS-Bench: A Benchmark for Generating Information Delivery Specification from BIM Information Requirements
- Subliminal Steering: Stronger Encoding of Hidden Signals
- MedStruct-S: A Benchmark for Key Discovery, Key-Conditioned QA and Semi-Structured Extraction from OCR Clinical Reports
- The End of Transformers? On Challenging Attention and the Rise of Sub-Quadratic Architectures
- Untangling the Mechanisms of Misleading Context in Medical Question Answering