Search papers, labs, and topics across Lattice.
This paper introduces CoRT, a novel token-level credit weighting method designed to enhance rubric-based reinforcement learning for language models by addressing the limitations of scalar response-level rewards. By employing counterfactual replay to generate tokenwise log-likelihood contrasts, CoRT effectively redistributes the GRPO advantage across tokens based on their contextual relevance to the rubric. The method outperforms traditional response-level GRPO approaches, achieving an average improvement of 4.4 percentage points across various instruction-tuned models, while maintaining competitive performance against more complex learned token-level baselines.
Token-level credit allocation using counterfactual replay boosts performance by an average of 4.4 percentage points without the need for an auxiliary scoring model.
Rubric-based reinforcement learning enriches language model training by evaluating model outputs against explicit criteria. Yet in GRPO-style pipelines, these structured judgments are reduced to a scalar response-level reward and converted into a response-level advantage, which is broadcast uniformly to all generated tokens. This leaves no explicit mechanism for allocating credit within a response, even when different criteria are grounded in different spans, formatting decisions, or semantic choices. We propose CoRT, a token-level credit weighting method for rubric-conditioned GRPO. Instead of training an auxiliary token scoring model, CoRT uses counterfactual replay to rescore the same sampled response under the original rubric-conditioned prompt and a matched criteria-free prompt. The resulting tokenwise log-likelihood contrasts serve as a proxy for dependence on the rubric context. CoRT maps these contrasts to bounded, response-normalized weights and uses them to redistribute the signed GRPO advantage across tokens, without introducing an auxiliary scorer or changing the response-level reward. Experiments across instruction-tuned models and reward granularities show that CoRT improves over matched response-level GRPO in the vast majority of comparisons, with an average gain of 4.4 percentage points. The method remains competitive with learned token-level credit baselines while avoiding a separate relevance-learning stage. These results suggest that policy-internal counterfactual likelihood contrasts provide an effective training signal for within-response credit allocation while retaining the simplicity and stability of GRPO.
