Rlvr Reasoning Eval

---
name: rlvr-reasoning-eval
description: This evaluation probes the mathematical and out-of-domain reasoning capabilities of large language models trained with Reinforcement Learning with Verifiable Rewards (RLVR). It specifically tests how well entropy-aware credit assignment methods allocate learning signals across high-entropy tokens during chain-of-thought generation. Performance is measured by average accuracy and pass rate over multiple sampled reasoning paths. Use when the user wants to benchmark on AIME24, AIME25, AMC, MATH, Minerva, Olympiad, or asks about evaluating this task. Reports Avg@k, Pass@k.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2604.11056
  bibtex_key: he2026rethinking
  confidence: high
---

# rlvr-reasoning-eval

> Rethinking Token-Level Credit Assignment in RLVR: A Polarity-Entropy Analysis — He et al. (2026) (arXiv:2604.11056, 2026)

## What this evaluates

This evaluation probes the mathematical and out-of-domain reasoning capabilities of large language models trained with Reinforcement Learning with Verifiable Rewards (RLVR). It specifically tests how well entropy-aware credit assignment methods allocate learning signals across high-entropy tokens during chain-of-thought generation. Performance is measured by average accuracy and pass rate over multiple sampled reasoning paths.

## Datasets

- **AIME24** — total ?; splits: test (-1)
- **AIME25** — total ?; splits: test (-1)
- **AMC** — total ?; splits: test (-1)
- **MATH** — total ?; splits: test (-1)
- **Minerva** — total ?; splits: test (-1)
- **Olympiad** — total ?; splits: test (-1)

## Metrics

- `Avg@k` **(primary)** — range: percent
  - Average accuracy across k sampled responses: (1/k) * Σ I(response_i is correct).
- `Pass@k` **(primary)** — range: percent
  - Probability that at least one of k sampled responses is correct: 1 - (1/k) * Σ I(response_i is incorrect).

## Input / output format

**Input**: Mathematical reasoning problems from the listed benchmarks.

**Output**: Autoregressive chain-of-thought reasoning followed by a final answer.

## Scoring recipe

```python
def compute_metrics(k, predictions, gold):
    correct = [1 if pred == gold else 0 for pred in predictions]
    avg_k = sum(correct) / k
    pass_k = 1.0 if any(correct) else 0.0
    return avg_k, pass_k
```

## Common pitfalls

- Different sampling counts (k=32 vs k=4) are used across benchmarks, so raw scores are not directly comparable without noting the k value.
- Stochastic decoding variance is mitigated by averaging over multiple samples, but results may still fluctuate depending on temperature and top-p settings during inference.
- Evaluation temperature (0.6) differs from training rollout temperature (1.0), which can shift the distribution of generated reasoning paths.

## Evidence (verbatim from paper)

> We evaluate models on six widely used reasoning benchmarks: AIME24, AIME25, AMC, MATH, Minerva, and Olympiad. We use vLLM with a temperature 0.6 and top-p 0.95 for inference. To reduce variance in stochastic decoding, we report Avg@32 and Pass@32 for AIME24, AIME25, and AMC, and Avg@4 and Pass@4 for the remaining benchmarks.

## Citation

```bibtex
@misc{he2026rethinking,
  title={Rethinking Token-Level Credit Assignment in RLVR: A Polarity-Entropy Analysis},
  author={He et al. (2026)},
  year={2026},
  note={arXiv:2604.11056}
}
```

- arXiv: 2604.11056