Metraipolis › archive › dc_956df3d6adcb

Hallucination Mitigation in Large Language Models: A Review and Exploratory Small-Model Pilot (2026) · DOI 10.20944/preprints202609.1923.v1

Source
Metraipolis archive
Original
https://doi.org/10.20944/preprints202609.1923.v1
License (as archived)
CC-BY-4.0
Archived
2026-10-08T01:19:23.331Z
SHA-256 of the archived text
d75a368fccde327ca9297f167f9f4c1f19fe54a21351a54a3458e2940ebb5340
Stable URL of this page
https://metraipolis.com/documento/dc_956df3d6adcb

How to cite

Hallucination Mitigation in Large Language Models: A Review and Exploratory Small-Model Pilot (2026) · DOI 10.20944/preprints202609.1923.v1 — Metraipolis archive, https://doi.org/10.20944/preprints202609.1923.v1 (CC-BY-4.0); archived copy: https://metraipolis.com/documento/dc_956df3d6adcb (Metraipolis archive dc_956df3d6adcb, sha256 d75a368fccde327ca9297f167f9f4c1f19fe54a21351a54a3458e2940ebb5340)
BibTeX
@misc{dc956df3d6adcb,
  title = {Hallucination Mitigation in Large Language Models: A Review and Exploratory Small-Model Pilot (2026) · DOI 10.20944/preprints202609.1923.v1},
  howpublished = {Metraipolis archive},
  url = {https://doi.org/10.20944/preprints202609.1923.v1},
  note = {License CC-BY-4.0; archived copy https://metraipolis.com/documento/dc_956df3d6adcb; sha256 d75a368fccde327ca9297f167f9f4c1f19fe54a21351a54a3458e2940ebb5340},
  year = {2026}
}

Archived text

Reproduced under CC-BY-4.0. Attribution: Metraipolis archive, https://doi.org/10.20944/preprints202609.1923.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: Hallucination Mitigation in Large Language Models: A Review and Exploratory Small-Model Pilot
Autori: Ali Ömer Yasan
Pubblicato: 2026-09-22 · editore: MDPI AG
DOI: https://doi.org/10.20944/preprints202609.1923.v1
Licenza dell'articolo: CC-BY-4.0

Abstract (testo degli autori, CC-BY-4.0):
Large language models (LLMs) can produce fluent responses that are false, unsupported, outdated, or more certain than the available evidence justifies. This manuscript presents a narrative review of major hallucination-mitigation approaches and an exploratory pilot intended to refine a future evaluation protocol rather than establish model-level performance. The review covers false-premise handling, abstention and knowledge-boundary estimation, retrieval-augmented generation, grounding, verification, self-checking, and multimodal hallucination. It also distinguishes published evidence from an engineering proposal for failure-driven correction. The pilot used a locally executed interface labeled Gemma 3 1B-IT and ten hand-designed stress prompts spanning false premises, insufficient evidence, fabricated-citation traps, unknowable events, and future-value requests. The revised analysis deliberately avoids compressing these observations into a single hallucination score. Two responses were epistemically appropriate, three were partially appropriate but introduced unsupported content, and five failed the target behavior. Generation stability was tracked separately: one otherwise appropriate answer entered a repetition loop, and another item required a restart after a context-length error. An extended follow-up run was terminated when the local system became unstable and began returning empty outputs; those observations are reported only qualitatively. The pilot is too small and insufficiently controlled to support causal claims or generalization to the Gemma family. Its value is methodological: it illustrates why hallucination research should report multiple axes - factual correctness, false-premise resistance, correct abstention, unnecessary abstention, grounding fidelity, coverage, calibration where available, and generation stability - rather than one aggregate number. The literature further suggests that mitigation is most credible when model behavior, external evidence, verification, and evaluation design are treated as a coupled system.

Fonte: https://api.crossref.org/works/10.20944%2Fpreprints202609.1923.v1 — metadati Crossref, consultati il 2026-10-07. Interrogazione: «small language models».