External factual authority
Wikidata remains explicit, inspectable, replaceable, and accessed through deterministic tooling. Learned weights are not treated as the source of factual truth.
A conservative path from fallback experience to local capability
Large language model cascades reduce inference cost by routing easy requests to smaller models, and distillation can further improve those models over time. We study a stricter neuro-symbolic variant: declarative facts remain external and replaceable, while only procedural skills migrate into learned parameters. We call this process verified skill migration (VSM).
A compact local model attempts a procedural subtask, a frozen competence gate escalates uncertain cases to an LLM, and the LLM outcome becomes a training trace only after independent verification using withheld ground truth or deterministic knowledge-graph evidence. Tiny instantiates VSM for Wikidata QA with entity linking and property mapping as migratable skills and a pinned local Wikidata snapshot as factual authority.
The model learns procedures, not the factual store
Verified migration
A teacher proposal becomes training data only after independent outcome verification. Frozen competence gates decide whether a compact skill may act autonomously.
Verified Skill Migration
UNKNOWN
FALLBACK TO LLM
VERIFY OUTCOME
STORE TRACE
TRAIN SKILL
FROZEN GATE
PROMOTE
FEWER FALLBACKS
A visible authority boundary
Question
Compact Skill
Frozen Competence Gate
Deterministic KG / runtimeExternal factual answer
Large LLMVerificationVerified TraceTraining → Compact Skill ↑
Fewer real fallback calls, broader high-precision autonomy
Scope: LC-QuAD 2.0 official TEST, preregistered structural subset, n = 1,074. This is not a full LC-QuAD 2.0 QA benchmark result.
207real Qwen calls saved
17.1%relative reduction in calls / question
96.51%learned zero-call precision
| Metric | Baseline | Learned |
|---|---|---|
| Real Qwen calls | 1,214 | 1,007 |
| Calls / question | 1.130 | 0.938 |
| Entity hybrid accuracy | 72.53% | 74.02% |
| Property PID accuracy | 66.57% | 68.99% |
| Exact (entity, property) accuracy | 63.59% | 65.64% |
| Zero-call coverage | 25.98% | 34.64% |
| Zero-call precision | 93.91% | 96.51% |
Fallback displacement is the robust result
Call reduction
Calls per question decreased from 1.130 to 0.938. The paired bootstrap 95% confidence interval for the absolute reduction is [0.153, 0.233] calls per question.
Accuracy
Exact pair accuracy increased from 63.59% to 65.64%, but the paired confidence interval includes zero and McNemar p = 0.107. The paper does not claim a statistically significant accuracy improvement.
Central claim. Verified Skill Migration substantially reduces real LLM fallback usage and expands a high-precision autonomous region, without evidence that the frozen accuracy non-inferiority requirement is violated.
Promotion is earned, not assumed
| Skill | Result |
|---|---|
| Entity linking | PASS |
| Property mapping | PASS |
| Relation directionality | FAIL / not promoted |
| Answer-type prediction | FAIL / not promoted |
The negative studies are preserved in the paper. They show that VSM does not claim every candidate skill can be transferred automatically from an LLM.
Frozen evidence with trace-level provenance
Included evidence
Formal result, statistical analysis, paired baseline and learned traces, executed v1.56.2 script snapshot, paper sources, and the original SHA-256 publication manifest.
Guardrail
The LC-QuAD 2.0 TEST result is frozen. No new threshold tuning, prompt optimization, checkpoint selection, or rescue runs should be performed on this result.
Cite Tiny
Version 1.0.0 is archived at Zenodo as 10.5281/zenodo.22195009. The permanent arXiv identifier will be added after announcement.
@article{revyakin2026tiny,
title = {Tiny: Verified Skill Migration for Reducing LLM Fallbacks
in Neuro-Symbolic Question Answering},
author = {Revyakin, Alexander},
year = {2026},
doi = {10.5281/zenodo.22195009},
note = {Version 1.0.0; arXiv identifier pending}
}
Version 1.0.0
The canonical repository location is github.com/alexrev1989/tiny-verified-skill-migration. Version 1.0.0 is archived at Zenodo under DOI 10.5281/zenodo.22195009. The arXiv submission is awaiting announcement.
Scientific content and the LC-QuAD 2.0 TEST result remain frozen.