Metraipolis › archive › dc_39610d49ac45

Policy-TD: Teacher-Distilled Runtime Control for Frozen Small Language Models (2026) · DOI 10.21203/rs.3.rs-10846457/v1

Source
Metraipolis archive
Original
https://doi.org/10.21203/rs.3.rs-10846457/v1
License (as archived)
CC-BY-4.0
Archived
2026-10-08T01:20:43.421Z
SHA-256 of the archived text
76aad8dd2764b908bbab9807ef599164714758b4a86c03057e310b104f628044
Stable URL of this page
https://metraipolis.com/documento/dc_39610d49ac45

How to cite

Policy-TD: Teacher-Distilled Runtime Control for Frozen Small Language Models (2026) · DOI 10.21203/rs.3.rs-10846457/v1 — Metraipolis archive, https://doi.org/10.21203/rs.3.rs-10846457/v1 (CC-BY-4.0); archived copy: https://metraipolis.com/documento/dc_39610d49ac45 (Metraipolis archive dc_39610d49ac45, sha256 76aad8dd2764b908bbab9807ef599164714758b4a86c03057e310b104f628044)
BibTeX
@misc{dc39610d49ac45,
  title = {Policy-TD: Teacher-Distilled Runtime Control for Frozen Small Language Models (2026) · DOI 10.21203/rs.3.rs-10846457/v1},
  howpublished = {Metraipolis archive},
  url = {https://doi.org/10.21203/rs.3.rs-10846457/v1},
  note = {License CC-BY-4.0; archived copy https://metraipolis.com/documento/dc_39610d49ac45; sha256 76aad8dd2764b908bbab9807ef599164714758b4a86c03057e310b104f628044},
  year = {2026}
}

Archived text

Reproduced under CC-BY-4.0. Attribution: Metraipolis archive, https://doi.org/10.21203/rs.3.rs-10846457/v1. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Articolo recente con licenza aperta: scheda dai metadati Crossref (posted-content).
Titolo: Policy-TD: Teacher-Distilled Runtime Control for Frozen Small Language Models
Autori: Surya Teja Avvaru, Krishna Sai Pokala, VARUN KUMAR REDDY DODDA
Pubblicato: 2026-09-01 · editore: Springer Science and Business Media LLC
DOI: https://doi.org/10.21203/rs.3.rs-10846457/v1
Licenza dell'articolo: CC-BY-4.0

Abstract (testo degli autori, CC-BY-4.0):
Abstract Small language models are attractive for local, low-cost, and privacy-sensitive deployment, but an important class of failures occurs when a model commits to an answer despite insufficient or contradictory evidence. We study this as a runtime-control problem. Policy-TD leaves student weights frozen and distills teacher-audited trace states into a compact multi-head controller that predicts semantic diagnostics, an intervention gate, and one of six bounded actions: finalization, abstention, repair, rollback, grounded continuation, or ordinary continuation. Across three independently trained guide seeds and three frozen student models, internal heldout accuracy rises from 31.0% to 47.1–47.3%, with 145–147 helped cases and zero observed harmed cases; router-off returns performance to 31.0%. External-style stress accuracy rises from 33.56% to 37.63% with 110 helped and zero observed harmed cases. Public external accuracy changes only from 18.67% to 19.22%; under the conservative setting, 26 cases are helped and 6 harmed. Among conservative behavior-changing interventions, help yield falls from 74.0% internally to 4.8% on public external evaluation, showing that transfer is limited by recoverability as well as by intervention targeting. Gains concentrate in missing-information and unknowable-answer families. The evidence therefore supports a narrow claim: decision-space adaptation can recover answer-commitment failures when the frozen generator retains recoverable competence, but it does not create missing model capability.

Fonte: https://api.crossref.org/works/10.21203%2Frs.3.rs-10846457%2Fv1 — metadati Crossref, consultati il 2026-10-07. Interrogazione: «small language models».