Ttsr Eval

ttsr-eval

TTSR: Test-Time Self-Reflection for Continual Reasoning Improvement — He et al. (2026) (arXiv:2603.03297, 2026)

What this evaluates

Evaluates the ability of LLMs to improve reasoning performance at test time through self-reflection and targeted variant question synthesis, without external supervision. It probes how well a model can adapt its policy to difficult mathematical and general reasoning problems by diagnosing its own failures and generating corrective training signals.

Datasets

• AMC23 — total ?; splits: test (-1)
• MATH-500 — total ?; splits: test (-1)
• Minerva — total ?; splits: test (-1)
• OlympiadBench — total ?; splits: test (-1)
• AIME 2024 — total ?; splits: test (-1)
• AIME 2025 — total ?; splits: test (-1)
• GPQA-Diamond — total ?; splits: test (-1)
• MMLU-Pro — total ?; splits: test (-1)

Metrics

• accuracy (primary) — range: percent
  • Percentage of test instances where the model's final extracted answer exactly matches the ground truth label.

Input / output format

Input: A mathematical or general reasoning problem statement.

Output: A step-by-step reasoning trajectory followed by a final answer.

Scoring recipe

def compute_accuracy(predictions, gold_answers):
    correct = sum(1 for p, g in zip(predictions, gold_answers) if p.strip() == g.strip())
    return (correct / len(predictions)) * 100

Common pitfalls

• Test-time updates may overfit to the specific adaptation instances rather than learning reusable reasoning patterns, leading to poor cross-dataset generalization.
• Self-consistency based pseudo-labels can be highly noisy when the majority of sampled trajectories are incorrect, potentially destabilizing the policy update if not properly regularized.

Evidence (verbatim from paper)

We report accuracy (%) or the official score for each benchmark.

Citation

@misc{he2026ttsr,
  title={TTSR: Test-Time Self-Reflection for Continual Reasoning Improvement},
  author={He et al. (2026)},
  year={2026},
  note={arXiv:2603.03297}
}

• arXiv: 2603.03297