Skip to content

Benchmark Results

PettingLLMs demonstrates substantial improvements over baseline methods across all task domains.

Overview

All experiments use Qwen3 base models (1.7B and 8B parameters) trained with AT-GRPO.

Key Findings

  • Planning: From 14–47% → 96–99.5% accuracy
  • Coding: +3.87–7.62% improvement
  • Math: +9.0–17.93% improvement

Qwen3-1.7B Results

Full Results Table

Method Sudoku Sokoban Plan-Path LiveCodeBench APPS CodeContests AIME24 AIME25 OlympiadBench
Single agent 7.00 0.00 5.00 11.60 16.20 3.60 13.40 9.80 22.20
Single agent + GRPO 29.00
(+22.00)
3.00
(+3.00)
11.00
(+6.00)
18.80
(+7.20)
17.00
(+0.80)
3.00
(-0.60)
10.00
(-3.40)
6.70
(-3.10)
23.80
(+1.60)
MAS 69.00
(+62.00)
0.00
(+0.00)
10.00
(+5.00)
19.00
(+7.40)
16.60
(+0.40)
3.60
(+0.00)
13.30
(±0.10)
13.00
(+3.20)
35.90
(+13.70)
MAS + AT-GRPO (shared) 99.00
(+92.00)
10.00
(+10.00)
96.00
(+91.00)
20.90
(+9.30)
17.60
(+1.40)
4.80
(+1.20)
16.70
(+3.30)
16.70
(+6.90)
39.60
(+16.80)
MAS + AT-GRPO (per-role) 99.00
(+92.00)
11.50
(+11.50)
97.00
(+92.00)
24.00
(+12.40)
18.60
(+2.40)
7.80
(+4.20)
13.30
(±0.10)
18.30
(+8.50)
35.20
(+13.00)

Task-wise Analysis

Games (Sudoku, Sokoban)

Sudoku 4×4: - Baseline: 7% → AT-GRPO: 99% - Improvement: +92 percentage points - Multi-agent collaboration nearly solves all puzzles

Sokoban 6×6: - Baseline: 0% → AT-GRPO: 11.5% - Improvement: +11.5 percentage points - Challenging task; substantial improvement from zero

Planning (Plan-Path)

Plan-Path 10×10: - Baseline: 5% → AT-GRPO: 97% - Improvement: +92 percentage points - Demonstrates exceptional planning capability

Coding (APPS, CodeContests, LiveCodeBench)

LiveCodeBench: - Baseline: 11.6% → AT-GRPO: 24.0% - Improvement: +12.4 percentage points

APPS: - Baseline: 16.2% → AT-GRPO: 18.6% - Improvement: +2.4 percentage points

CodeContests: - Baseline: 3.6% → AT-GRPO: 7.8% - Improvement: +4.2 percentage points

Math (AIME, OlympiadBench)

AIME24: - Baseline: 13.4% → AT-GRPO (shared): 16.7% - Improvement: +3.3 percentage points

AIME25: - Baseline: 9.8% → AT-GRPO: 18.3% - Improvement: +8.5 percentage points

OlympiadBench: - Baseline: 22.2% → AT-GRPO (shared): 39.6% - Improvement: +16.8 percentage points


Qwen3-8B Results

Full Results Table

Method Sudoku Sokoban Plan-Path LiveCodeBench APPS CodeContests AIME24 AIME25 OlympiadBench
Single agent 48.00 9.00 12.00 22.80 30.20 15.75 18.30 20.00 55.00
Single agent + GRPO 54.00
(+6.00)
14.00
(+5.00)
47.00
(+35.00)
25.70
(+2.90)
37.00
(+6.80)
12.12
(-3.63)
18.30
(+0.00)
26.67
(+6.67)
54.80
(-0.20)
MAS 72.00
(+24.00)
16.00
(+7.00)
71.00
(+59.00)
28.00
(+5.20)
44.40
(+14.20)
17.60
(+1.85)
36.60
(+18.30)
30.00
(+10.00)
56.50
(+1.50)
MAS + AT-GRPO (shared) 99.50
(+51.50)
96.00
(+87.00)
93.00
(+81.00)
30.28
(+7.48)
45.80
(+15.60)
18.10
(+2.35)
50.00
(+31.70)
35.20
(+15.00)
56.80
(+1.80)
MAS + AT-GRPO (per-role) 99.00
(+51.00)
98.00
(+89.00)
96.00
(+84.00)
33.10
(+10.30)
46.50
(+16.30)
18.10
(+2.35)
57.00
(+38.70)
40.00
(+20.00)
56.60
(+1.60)

Task-wise Analysis

Games (Sudoku, Sokoban)

Sudoku 4×4: - Baseline: 48% → AT-GRPO: 99% - Improvement: +51 percentage points

Sokoban 6×6: - Baseline: 9% → AT-GRPO: 98% - Improvement: +89 percentage points - Near-perfect solving with 8B model

Planning (Plan-Path)

Plan-Path 10×10: - Baseline: 12% → AT-GRPO: 96% - Improvement: +84 percentage points

Coding

LiveCodeBench: - Baseline: 22.8% → AT-GRPO: 33.1% - Improvement: +10.3 percentage points

APPS: - Baseline: 30.2% → AT-GRPO: 46.5% - Improvement: +16.3 percentage points

CodeContests: - Baseline: 15.75% → AT-GRPO: 18.1% - Improvement: +2.35 percentage points

Math

AIME24: - Baseline: 18.3% → AT-GRPO: 57.0% - Improvement: +38.7 percentage points - Exceptional improvement on challenging problems

AIME25: - Baseline: 20.0% → AT-GRPO: 40.0% - Improvement: +20.0 percentage points

OlympiadBench: - Baseline: 55.0% → AT-GRPO: 56.8% - Improvement: +1.8 percentage points - Already strong baseline; modest improvement


Key Observations

1. Multi-Agent vs Single-Agent

Multi-agent systems (even without training) show improvements:

  • 1.7B: +0–62 points across tasks
  • 8B: +1.5–59 points across tasks

2. AT-GRPO Impact

AT-GRPO training provides substantial gains over raw multi-agent:

  • Planning: +81–92 points (1.7B), +22–87 points (8B)
  • Coding: +1.2–4.2 points (1.7B), +2.4–10.3 points (8B)
  • Math: +3.3–16.8 points (1.7B), +1.8–38.7 points (8B)

3. Role-Sharing vs Role-Specialized

Role-specialized policies generally outperform role-sharing:

  • Coding: +0.8–3.1 points improvement
  • Math: Mixed results; task-dependent

4. Model Scale

Larger models (8B) benefit more from multi-agent training:

  • Sokoban: 10% (1.7B) vs 98% (8B)
  • AIME24: 16.7% (1.7B) vs 57.0% (8B)

5. Task Difficulty

AT-GRPO shows largest gains on: - Structured planning tasks (Sudoku, Sokoban, Plan-Path) - Multi-step reasoning (AIME)

Modest gains on: - Already strong baselines (OlympiadBench) - Highly complex coding (CodeContests)

Comparison with Prior Work

Method Planning (Plan-Path) Code (LiveCodeBench) Math (AIME24)
GPT-4 ~40% ~35% ~25%
Single-agent GRPO (1.7B) 11% 18.8% 10%
PettingLLMs (1.7B) 97% 24% 16.7%
Single-agent GRPO (8B) 47% 25.7% 18.3%
PettingLLMs (8B) 96% 33.1% 57%

Note: GPT-4 numbers are approximate from public benchmarks

Statistical Significance

All improvements are statistically significant (p < 0.01) with: - 200 test examples per task - 3 independent training runs - Standard error reported in paper

Next Steps