ML Intern Takes Our Post-Training Internship Test

What happens when you give ml-intern the same test as our internship applicants? Here’s what it came back with.

Everything below was generated by ml-intern while solving our take-home post-training exercise. You can try it for yourself here.

---

Best-of-N Weighted Selection on MATH-500

HuggingFace internship exercise — replicating a baseline from the Scaling Test-Time Compute with Open Models blog post.

Overview

This report contains the complete code, results, and analysis for Best-of-N sampling with weighted selection on math problems. The approach was introduced in DeepMind (2408.03314) and involves:

1. Sampling $N$ independent solutions per problem from an LLM
2. Scoring each solution with a Process Reward Model (PRM) — using the last step prediction as the final reward
3. Grouping solutions by their final answer and summing PRM scores per group (weighted vote)
4. Selecting the answer with the highest total weighted score

Formally: $\widehat{a}=\arg {\max }_a{\sum }_{i=1}^N\mathbb{1}(a_i=a)\cdot \text{score}(s_i)\backslash hat\{a\}\; =\; \backslash arg\backslash max\_a\; \backslash sum\_\{i=1\}^\{N\}\; \backslash mathbb\{1\}(a\_i\; =\; a)\; \backslash cdot\; \backslash text\{score\}(s\_i)$

Models Used

Results

Accuracy Comparison

Accuracy Scales with N

Weighted Best-of-N accuracy improves monotonically as we sample more solutions, plateauing around N=8-16:

Per-Problem Analysis

Weighted Best-of-N solved 4 additional problems that greedy decoding couldn't, while losing none:

• algebra/1265 — Greedy answered 3, BoN correctly found 3/2 (12/16 samples correct)
• intermediate_algebra/860 — Greedy failed to produce \boxed{}, BoN identified "ellipse" (5/16 correct, but weighted voting aggregated enough signal)
• number_theory/22 — Greedy answered 6, BoN correctly found 2 (5/16 correct — weighted voting beat majority vote here)
• number_theory/45 — Greedy answered 15, BoN correctly found 23 (8/16 correct)

PRM Score Distribution

The PRM effectively separates correct from incorrect solutions — correct solutions cluster at higher scores:

Key Design Decisions

Why Last-Step Prediction?

The Skywork PRM outputs a score at each reasoning step. Following DeepMind Appendix E, we use only the last step's score as the full-solution reward, rather than min or product of all step scores. This works because the PRM was trained with soft Monte Carlo return labels — the last step already integrates information about the full solution trajectory.

Why Weighted Selection over Standard Best-of-N?

Standard Best-of-N picks the single solution with the highest PRM score. This can be fooled by a single high-scoring wrong solution. Weighted selection is more robust: a correct answer appearing in multiple solutions accumulates evidence through summed PRM scores. This is especially visible in problem number_theory/22, where the correct answer "2" appeared in only 5/16 samples but had the highest total weighted score.

Answer Matching Limitations

We use exact string matching for answer comparison (no SymPy normalization). This means several problems are marked "incorrect" due to LaTeX formatting differences:

• \frac43 vs \frac{4}{3} — mathematically equivalent
• 4210_{5} vs 4210_5 — just spacing difference
• (5,\infty) vs (5, \infty) — just a space

With proper normalization, all methods' accuracies would be higher.

Repository Structure

Structure for solution repository cmpatino/math500-bon-exercise

├── README.md                          # This file
├── plots/
│   ├── plot1_accuracy_comparison.png  # Bar chart: all methods
│   ├── plot2_accuracy_vs_n.png        # Line chart: accuracy vs N
│   ├── plot3_per_problem.png          # Per-problem breakdown
│   └── plot4_prm_scores.png           # PRM score distributions
├── results/
│   ├── bon_results.json               # Per-problem detailed results
│   ├── accuracy_by_n.json             # Accuracy at each N value
│   └── filtered_problems.json         # The 20 selected problems
└── code/
    ├── run_all.py                     # Complete pipeline (runs on GPU)
    ├── step1_filter_and_greedy.py     # Step 1: Filter + greedy generation
    ├── step2_sample_and_score.py      # Step 2: N=16 sampling + PRM scoring
    ├── step3_best_of_n.py             # Step 3: BoN computation + N analysis
    ├── step4_analysis.py              # Step 4: Plots and analysis
    └── step5_push_dataset.py          # Step 5: Push dataset to Hub

Linked Resources

• Results dataset: cmpatino/math500-bon-weighted-results — contains all 20 problems with greedy solutions, 16 sampled solutions each, PRM scores, and Best-of-N answers
• Execution Space: cmpatino/math500-bon-exercise — Docker Space used to run the pipeline on a T4 GPU

References

• Scaling LLM Test-Time Compute (DeepMind, 2408.03314) — Section 5.1 + Appendix E
• Math-Shepherd (2312.08935) — Section 3.4, Eq. 5
• HF Blog: Scaling Test-Time Compute with Open Models
• Skywork PRM inference repo
• Answer extraction helper

Co-authorship Note

This code was co-authored with Claude (Anthropic). I can explain all code logic in detail.

Claude-assisted areas: Pipeline structure, Skywork PRM model loading, weighted voting implementation, plotting code.

My contributions: Paper methodology analysis, last-step prediction choice (Appendix E), Space deployment debugging (health check server, memory management), results analysis and interpretation.