MÉTIS
Can a small, self-trained model help a larger local language model solve structured tasks more efficiently – without reducing correctness?
Research archive · P2 · experimentalLearned proposals instead of more LLM work?
MÉTIS is a small, specialized Transformer. It generates candidate solutions for already formalized constraint-solving tasks. Candidates can be checked independently and supplied as hints to a larger local language model.
The research question is whether learned hints provide a measurable benefit over no hint, random candidates, or deterministic heuristics – including the full cost of generating and verifying those hints.
P2: 679,571 parameters
The documented P2 architecture has 679,571 trainable parameters. The repository archives P2 source code, seven selected checkpoints, seven synthetic dataset splits and historical evaluations. MÉTIS is not a general-purpose language model; the recorded tasks were formalized in advance.
The archive offers integrity checks and saved-result reaggregation. Original training and historical Qwen executions have not been independently rerun for this documentation. Complete earlier P1 comparisons require some missing historical artifacts.
What is actually verifiable
Verifying archived checksums and recomputing stored summaries is not the same as independently repeating the historical experiments.
Four arms, twelve tasks
The archived comparison used qwen3.5:latest on twelve already formalized constraint tasks, at temperature 0 with thinking disabled. The arms were: no proposal, MÉTIS P2, a random candidate, and Min-Conflicts with 256 checks.
The published figures are a descriptive assembly of resumed partial runs (run_03 to run_05), not one preregistered, independently replicated experiment.
| Condition | Correct | Qwen output tokens | Qwen response time |
|---|---|---|---|
| A · No proposal | 11/12 | 40,559 | 2,045.3 s |
| B · MÉTIS P2 | 10/12 | 36,241 | 1,892.8 s |
| C · Random candidate | 10/12 | 52,074 | 2,819.2 s |
| D · Min-Conflicts (256 checks) | 11/12 | 24,910 | 1,367.5 s |
Interpretation: MÉTIS P2 used 10.6% fewer Qwen output tokens than A but solved one fewer task. Min-Conflicts used the fewest tokens while matching A's correct count. The reported times measure Qwen responses, not all additional end-to-end costs.
The snapshot therefore does not establish a general acceleration effect or a causal advantage from learned MÉTIS hints. Twelve tasks are insufficient for broad generalization.
Reported, not independently reproduced
The project report dated 18 October 2026 describes further training. For a learned control policy (C4), it reports utility +0.015 versus a rule at p < 0.001. P2 did not demonstrate this benefit (p = 0.906).
The repository currently contains no P3 checkpoints, complete raw evaluations, or full training configuration to independently reconstruct these values. Reported performance on hard K4 tasks still trails Min-Conflicts. The Qwen comparison above used P2, not P3.
What remains unknown
Whether MÉTIS can speed up a larger model on constrained hardware under a full end-to-end cost comparison remains unknown. Controlled repeat trials would need pinned model digests, matched tasks and compute budgets, and explicit accounting for proposal generation, verification, loading, and caching.
The archive notes the risk of incorrect anchoring hints that can reduce correctness. Results have not demonstrated general natural-language reasoning or broad improvements to language models.
The repository is now public, but its release checklist still contains open items, including provenance, redistribution rights, model artifacts and independent review. Under the repository's licensing note, Apache-2.0 covers eligible original source code and documentation, not automatically all archived weights or datasets.
Evidence rather than marketing
Reducing output tokens is insufficient if fewer tasks are solved or costs are excluded. Negative results and strong deterministic baselines are part of the research record.
Read methods and evidence
The following documents are now available in the public GitHub repository. The links are pinned to the reviewed commit for traceability.
- README – Research overview
- Methods and comparison settings
- Findings and four-arm comparison
- P3 – Reported progress
- Research limitations
- Reproducibility
- Licensing scope and rights
- Outstanding release checks
Source checked: 20 October 2026 · GitHub commit 0c8ce163. This is a comparison against the published research documentation, not an independent replication of the experiments.