Skip to content
OpenTrain AIFor AI Companies

HFEPX · Eval paper review

Calibrated Decision Rewards: Training Language Models with a Judge That Knows How Sure It Is

Rudransh Agnihotri, Manasvi Kapoor

Published

Nov 18, 2026

Citations

0

Trust level

Moderate

Usefulness score

37/100 (Low)

Extraction confidence

55% (Moderate)

Derived from extracted protocol signals and abstract evidence.

Rater population

Not reported

Signals refreshed

Nov 18, 2026

Should you rely on this paper?

This paper is adjacent to HFEPX scope and is best used for background context, not as a primary protocol reference.

Use this for comparison and orientation, not as your only source.

Best use

Background context only

Use if you need

A benchmark-and-metrics comparison anchor.

What to verify

Validate the evaluation procedure and quality controls in the full paper before operational use.

Main weakness

No major weakness surfaced.

Human feedback signal
Not explicit
Not explicit in abstract metadata
Evaluation signal
Detected
Eval setup described
Usefulness for eval research
37/100
Adjacent candidate

Treat as adjacent context, not a core eval-method reference.

Abstract

Reinforcement learning (RL) can only teach a language model what its reward can recognize. Answer checkers are exact but need a reference answer; learned reward models and LLM judges work everywhere, but their scores are not probabilities. We study RLCDR (Reinforcement Learning from Calibrated Decision Rewards). An LLM converted into a decision model answers a typed question about a response ("is the final answer correct?", "which step is the first error?") with a probability distribution, read in one forward pass, debiased over option orders and calibrated with about a hundred labels. Training Qwen3-1.7B with GRPO on MATH, this reference-free reward raises MATH500 accuracy from 72.2% to 76.8% (+4.65 points, p < 1e-4), matching training on the answer key (76.9%), while a random reward gains only +0.65. The gain grows with difficulty, from 44.8% to 51.7% on the hardest level. On the policy's own changing samples the reward stays calibrated: its expected calibration error is half that of a softmax judge on the same 27B backbone (0.074 vs. 0.141), and it shows no sign of reward hacking. Without step training, the same model finds the first wrong step better than a dedicated 7B process reward model (ProcessBench F1 0.772 vs. 0.722, 20/20 splits). We also map where the approach stops being enough: trained only on problems the base model mostly fails, the reference-free reward still improves the policy (+1.9 MATH500 points, p = 0.03) but trails the answer key, because its probabilities become overconfident on that distribution. We state six hypotheses in advance and test each with a control designed to falsify it.

What we could verify

These are the protocol signals we could actually recover from the available paper metadata. Use them to decide whether this paper is worth deeper reading.

Human Feedback Types

missing

None explicit

No explicit feedback protocol extracted.

"Reinforcement learning (RL) can only teach a language model what its reward can recognize."

Evaluation Modes

strong

Llm As Judge, Automatic Metrics

Includes extracted eval setup.

"Reinforcement learning (RL) can only teach a language model what its reward can recognize."

Quality Controls

missing

Not reported

No explicit QC controls found.

"Reinforcement learning (RL) can only teach a language model what its reward can recognize."

Benchmarks / Datasets

strong

MATH

Useful for quick benchmark comparison.

"Training Qwen3-1.7B with GRPO on MATH, this reference-free reward raises MATH500 accuracy from 72.2% to 76.8% (+4.65 points, p < 1e-4), matching training on the answer key (76.9%), while a random reward gains only +0.65."

Reported Metrics

strong

Accuracy, F1

Useful for evaluation criteria comparison.

"Training Qwen3-1.7B with GRPO on MATH, this reference-free reward raises MATH500 accuracy from 72.2% to 76.8% (+4.65 points, p < 1e-4), matching training on the answer key (76.9%), while a random reward gains only +0.65."

Benchmarks and datasets

MATH

Reported metrics

accuracyf1
Human feedback details
Uses human feedback
No
Feedback types
None
Rater population
Not reported
Expertise required
Math
Evaluation details
Evaluation modes
Llm As Judge, Automatic Metrics
Agentic eval
None
Quality controls
Not reported
Evidence quality
Moderate
Use this page as
Background context only

Research brief

Metadata summary

Reinforcement learning (RL) can only teach a language model what its reward can recognize.

Based on abstract + metadata only. Check the source paper before making high-confidence protocol decisions.

Key takeaways

  • Reinforcement learning (RL) can only teach a language model what its reward can recognize.
  • Answer checkers are exact but need a reference answer; learned reward models and LLM judges work everywhere, but their scores are not probabilities.
  • We study RLCDR (Reinforcement Learning from Calibrated Decision Rewards).

Researcher actions

  • Compare this paper against nearby papers in the same arXiv category before using it for protocol decisions.
  • Validate inferred eval signals (Automatic metrics) against the full paper.
  • Use related-paper links to find stronger protocol-specific references.

Caveats

  • Generated from abstract + metadata only; no PDF parsing.
  • Signals below are heuristic and may miss details reported outside the abstract.

Recommended queries

Contribution summary

  • Reinforcement learning (RL) can only teach a language model what its reward can recognize.
  • Answer checkers are exact but need a reference answer; learned reward models and LLM judges work everywhere, but their scores are not probabilities.

Why it matters for eval

  • On the policy's own changing samples the reward stays calibrated: its expected calibration error is half that of a softmax judge on the same 27B backbone (0.074 vs.

Researcher checklist

  • Human feedback protocol is explicit

    No explicit human feedback protocol detected.

  • Evaluation mode is explicit

    Detected: Llm As Judge, Automatic Metrics

  • Quality control reporting appears

    No calibration/adjudication/IAA control explicitly detected.

  • Benchmark or dataset anchors are present

    Detected: MATH

  • Metric reporting is present

    Detected: accuracy, f1