PTRM: A 7M-Parameter Tiny Recursive Model Beats Giant LLM Ensembles at 0.0001x Cost via Test-Time Noise Exploration

Forum topic · 小凯 · 2026-06-03

Summary

PTRM (Probabilistic Tiny Recursive Model) extends the 7M-parameter Tiny Recursive Model (TRM) by injecting Gaussian noise into latent states at each recursive supervision step, running K parallel trajectories, and using the model's pre-trained Q head (originally built for adaptive computation time early stopping) as a verifier to pick the best answer. This converts deterministic recursion—which gets trapped in bad latent basins—into probabilistic exploration, adding a parallelizable width axis to test-time scaling alongside the existing depth axis. On Pencil Puzzle Bench, accuracy rises from 62.6% to 91.2% (+28.6pp), with Sudoku-Extreme improving from 87.4% to 98.75% and Maze-Hard from 83.8% to 86.73%. The best configuration reportedly outperforms an ensemble of seven frontier LLMs (55.1%) at under 0.0001x of its cost. Limitations include task-dependent optimal noise scales (σ from 0.1 to 1.0), weak Q-head reliability on Maze-Hard (pass@K 96% vs best-Q@K 85.17%), and modest gains on ARC-AGI-2 and Heyawake.

PTRM (Probabilistic Tiny Recursive Model) takes the 7M-parameter Tiny Recursive Model (TRM) and turns its deterministic latent recursion into probabilistic exploration—achieving accuracy jumps at a fraction of frontier LLM cost.

Background: TRM's Deterministic Trap

TRM is elegant by design: instead of generating token-level chains of thought, it recursively iterates a latent state. With only 7M parameters, it already beats frontier LLMs on Sudoku-Extreme.

The problem: deterministic recursion gets stuck in bad latent basins. The paper identifies three trajectory patterns:

Fast success: converges to the correct region within a few steps, Q value rising in sync
Delayed success: oscillates in the wrong region, then suddenly escapes to the correct one
Total failure: trapped in a wrong basin, never escaping

> "TRM has the capabilities to solve significantly more problems but is limited by its standard inference procedure."

PTRM's Core Innovations

1. Probabilistic Exploration: Turning Depth into Width + Depth

LLMs sample multiple CoT trajectories and pick the best. PTRM brings this idea into continuous latent space:

Inject Gaussian noise ε ~ N(0, σ²I) into the latent state at each depth recursion step
Run K independent trajectories in parallel
Use the model's existing Q head as a verifier and select the answer with the highest Q value

Crucially, the Q head was already trained. During training it drives ACT (Adaptive Computation Time) early stopping, but at inference it was previously discarded—huge waste. Correct trajectories score around +6, wrong ones around -6, an extremely sharp separation.

2. Where to Inject Noise: Every Step, Not Just Initialization

The paper compares noise strategies:

Noise only on the initial latent state: poor results
Langevin sampling (gradient + noise): negative result—the gradient adds nothing
PTRM's approach: inject noise at every supervision step — this is the effective escape mechanism

> "The gain came from the noise, not the gradient."

3. Two-Axis Test-Time Scaling

Depth axis (existing): more supervision steps D, but sequential with diminishing returns (D=16→48 only +3.4pp)
Width axis (new): more parallel trajectories K, fully parallelizable with huge gains (K=1→100 gives +28.6pp)

Results

| Benchmark | Baseline TRM | PTRM best-Q@K | Gain | |-----------|-------------|----------------|------| | Pencil Puzzle Bench | 62.6% | 91.2% | +28.6pp | | Sudoku-Extreme | 87.4% | 98.75% | +11.35pp | | Maze-Hard | 83.8% | 86.73% | +2.93pp |

Comparison with frontier LLMs (Pencil Puzzle Bench Golden Set):

| Method | Accuracy | Cost per correct answer | |--------|----------|------------------------| | claude-opus-4-6@thinking (direct) | 34.7% | $0.40 | | Ensemble of 7 strongest LLMs (perfect verifier) | 55.1% | $8.51 | | PTRM (7M params) | 91.2% | $0.001 |

7M parameters vs hundreds of billions; $0.001 vs $8.51; 91.2% vs 55.1%.

Technical Details and Limitations

Noise Scale Is Task-Dependent

Optimal σ varies widely across tasks:

Sudoku-Extreme: σ=0.1 nearly saturates (pass@K 99.3%)
Maze-Hard: needs σ≈1.0 (more complex latent space requires stronger perturbation)
ARC-AGI-2: σ≈0.6 is optimal; performance drops beyond that

Q-Head Verifier Limitations

On PPBench and Sudoku, best-Q@K is within 1pp of pass@K—the verifier is nearly perfect. But:

Maze-Hard: pass@K 96% vs best-Q@K 85.17% — the Q head is unreliable here
ARC-AGI-2: modest gains (7.36%→8.47%)
Heyawake: already at a plateau (85.7%→85.7%)

The Q head's verifier capability is strongly task-dependent; stronger verification mechanisms are needed going forward.

Key Takeaways

1. Latent basin trapping may be a universal bottleneck for recursive models — any deterministic iterative system can get stuck 2. Massive waste of training signal — what the Q head learns during training was underused at inference 3. Width scaling beats depth scaling in practice — parallelism delivers far more than sequential deepening 4. The "scale is everything" paradigm is being challenged — 7M parameters + smart test-time strategy > hundreds of billions of parameters + brute-force inference

Paper: https://arxiv.org/abs/2605.19943