Ali Uyar

Research / Repairing frozen models

SOAR: Recovering Operational Overlays from Shared Semantics in Frozen Language Models

Introduces SOAR, residual adapters built by subtracting a semantic scaffold in dense-delta space, recovering strict-JSON and exact-quotation behavior at direct-adapter level with essentially unchanged answer F1.

Published
29 March 2026
Author
Ali Uyar
DOI
10.5281/zenodo.19313240
Code
github.com/aliuyar1234/overlay-algebra
Read the paper (PDF)

Abstract

Post-training adapters often mix task semantics with operational output behavior, which makes behavior reuse hard to study cleanly. We consider a deliberately controlled version of that problem. Starting from filtered SQuAD 1.1 question answering, we retain only examples with exactly one gold support sentence, compile targets deterministically, and evaluate with parsers rather than an LLM judge. On this fixed QA substrate, we train a semantic scaffold adapter for plain answers and direct single-overlay adapters for two deployment-style output contracts: strict JSON and exact support quotation. We then construct Semantic-Operational Adapter Residuals (SOAR) by subtracting the scaffold in effective dense-delta space and recompressing the result.

On the corrected full-scale study, both overlays recover cleanly. For JSON, S + R_J matches the direct adapter on strict schema validity (0.9986 vs. 0.9986) with essentially unchanged answer F1 (0.8344 vs. 0.8346). For exact quotation, S + R_Q matches or slightly exceeds the direct adapter on quote exactness (0.9414 vs. 0.9413) and answer F1 (0.8122 vs. 0.8121). Under the same output contracts, prompt-only control on the scaffold is near zero on the active overlay metric in both cases, so the recovered behavior is not explained by prompting alone.

We present these results as evidence for partial operational recoverability, not universal modularity. The validated claim surface is intentionally narrow: the main contribution of this paper is a semantics-fixed protocol, a dense-delta residualization method, and full-scale single-overlay recovery for two operational behaviors with a deterministic, auditable evaluation pipeline.

Why it matters

Post-training adapters often mix task semantics with output behavior. Holding the QA task fixed while the output contract varies shows how much of a behavior such as strict JSON can be isolated as a residual on a shared semantic scaffold, with parsers rather than an LLM judge doing the scoring.

Cite

@misc{uyar2026soar,
  title     = {SOAR: Recovering Operational Overlays from Shared Semantics in Frozen Language Models},
  author    = {Uyar, Ali},
  year      = {2026},
  doi       = {10.5281/zenodo.19313240},
  publisher = {Zenodo},
  url       = {https://doi.org/10.5281/zenodo.19313240}
}
AU-7
Overlay leaf: residual ΔR = TSVD(ΔF − 1.0·ΔS), one at a timeBase sheet: scaffold S, LoRA rank 16 on o_proj and down_projReference sheets: direct adapters F_J, F_QS + R beside its referencePrompt-only S beside its referenceSOAR takes a scaffold adapter S trained for plain answers and a direct adapter F trained for each output contract on frozen Qwen2.5-7B-Instruct, subtracts β·S from F in dense-delta space and recompresses the remainder with truncated SVD into a residual R. Stacked on the same scaffold, S + R reaches the direct adapter on the contract metric: strict schema validity 0.9986 against 0.9986, and exact quotation 0.9414 against 0.9413, with answer F1 within 0.0002. Prompting the contract on S alone scores 0.0000 on both metrics, so the recovered behaviour is not explained by prompting alone. Every number is a deterministic, parser-based score with greedy decoding on 10,554 filtered SQuAD 1.1 test questions, with β = 1.0 selected on validation from {0.5, 0.75, 1.0, 1.25}.