Intrinsic Sequence-Likelihood Confidence in Retrieval-Dominated Extractive QA: Two Pre-Specified Negatives, and What They Do and Do Not Attribute
Gunwoo Lee, Changmin Sung, Sang-Hwan Gwak, Ina Kim, Ji-Young Choi, Kyong-Ha Lee
cs.CL · cs.IR · cs.LG
Quick verdict
Pre-specified negatives: intrinsic LLM confidence fails as distillation trigger and router in retrieval-dominated extractive QA; rigorous but setting-bound.
Summary
The paper evaluates whether intrinsic sequence-likelihood confidence from fine-tuned 7-9B open LLMs can serve as a control signal in extractive, source-grounded document QA. Two uses are tested under criteria committed to version control before execution: (1) a distillation trigger selecting queries for further adaptation via a composite of mode disagreement, retrieval-context similarity, and parametric confidence; and (2) a routing-and-abstention policy choosing retrieval, parametric answering, or abstention per query. Both fail: the trigger misses pre-specified criteria on all four families (LLaMA-3.1, Qwen2.5, Gemma-2, Mistral), and the routing pilot (LLaMA only) passes only E-1/E-3, failing E-2/E-4. The structural explanation is a ceiling: questions were generated from passages containing their answers, so retrieval alone recovers 92-99.8% of best-case combined accuracy, leaving 0.2-0.7% for any router. Signal diagnostics show AUROC 0.65-0.81 and ECE 0.48-0.72, insensitive to scalar recalibration and inconsistently improved by token temperature. Selector ablations on three adapted combinations show no detectable downstream benefit from the confidence term; on Gemma, removing it flips the selector from failing to passing. Limitations: two of five combinations are unadapted base models after guard rollback; one combination's registered verdict consumed signals from a known-defective prompt template; routing ran on a single model; adaptation effects are tiny (+0.01-0.03 F1). Code and data are released on Zenodo.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Under criteria committed to version control before the runs, a confidence-bearing distillation trigger fails its registered criteria on all four 7-9B model families, and a routing-and-abstention policy passes only two of four criteria in a single-model (LLaMA) pilot.
- Because questions were generated from the passages containing their answers, retrieval alone recovers 92-99.8% of best-case combined accuracy under every correctness criterion tested, leaving per-query routers at most 0.2-0.7% to compete for.
- Sequence-likelihood confidence is weak before and after adaptation (AUROC 0.65-0.81, ECE 0.48-0.72), invariant to scalar recalibration, and not consistently improved by token-level temperature rescaling.
- Pre-specified selector ablations on three adapted combinations show no statistically detectable downstream benefit from the confidence term on any of nine seeds; on Gemma, removing the term changes the selector from failing to passing both registered criteria.
- A pre-registered full-parameter fine-tuning control shows the signal does not improve, undercutting the objection that LoRA under-training explains the weakness.
Strengths
Credible pre-specification
Table S-9 records, for every registered experiment, the criteria-fixing commit, UTC timestamp, and first-run evidence (e.g., trigger criteria fixed at commit aac9c76 on 15 Jul 02:31, first prediction file 04:31); decision rules naming R-2 and F-2 as discriminating controls were fixed in advance, making the negatives hard to dismiss as selective reporting.
Disclosure of defects most papers bury
Two of five 'adapted' combinations are identity adapters after capability-guard rollback (Section 4); the LLaMA/SciTech held-out set predates a prompt-template fix leaving 45-word answers against two-word golds; an E-2 evaluation-code error initially reported a pass and was corrected to a fail on the record (Section 7); the KBI prior's executed polarity contradicts its registered rationale (S2).
Criterion-robustness analysis
Statistical hygiene
Attribution discipline
Weaknesses
Ceiling close to built-in by construction
Parametric mode near floor
Uneven coverage and unexecuted plans
Dataset construction not bit-reproducible
Unconventional thresholds and thin margins
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper asks whether intrinsic sequence-likelihood confidence of adapted 7-9B open LLMs can serve as a deployment control signal in extractive, source-grounded document QA. Two uses are tested against criteria committed to version control before the runs: (1) a distillation trigger selecting, via a percentile-ranked composite of mode disagreement, retrieval-context similarity, and parametric confidence, the queries for further LoRA adaptation; and (2) a routing-and-abstention policy choosing retrieval, parametric answering, or abstention per query. The claims are deliberately narrow negatives: the trigger fails its criteria on all four families (Table 1); the routing pilot (LLaMA only) passes E-1/E-3 but fails E-2/E-4; and the failures are explicitly not attributed wholesale to the signal. The setting characterization — retrieval recovers 92-99.8% of best-case combined accuracy — is the load-bearing result on which the routing verdict rests.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the main finding of this paper?
That in extractive QA where questions were generated from passages containing their answers, an LLM's own sequence-likelihood confidence fails as a pre-specified distillation trigger (on all four 7-9B model families) and as a routing/abstention policy (pilot on LLaMA, two of four criteria passed), because retrieval alone already recovers 92-99.8% of the best achievable accuracy.
Why does retrieval dominate in this setting?
By construction, each question was generated from the passage containing its answer, so a retriever that surfaces the right chunk surfaces the answer. The paper measures this as a ceiling: best-case routing, chosen with hindsight knowledge of the answers, beats always-retrieval by only about 0.2-0.7% of accuracy, so no per-query router can gain meaningfully.
Does this prove LLM confidence signals are useless?
No. The paper reports weak but nonzero discrimination (AUROC 0.65-0.81) and explicitly bounds its verdict: confidence cannot help routing here because there is no gain available, and cannot improve trigger selection measurably beyond what context similarity alone provides. The claim is setting- and mechanism-specific, not a general verdict on confidence.
Is the code and data available?
Yes. A code record (1,395 files, MD5-pinned) and a data record (45 files) are released as supplementary materials and deposited on Zenodo (DOIs 10.5281/zenodo.22710121 and 10.5281/zenodo.22721044), with commit-level provenance for every registered verdict; only the original QA generation via a served GPT alias is not bit-reproducible.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Text extraction omitted portions of the paper (parts of Section 4, the tail of Section 7, and fragments of supplementary tables S6/S19); judgments rest on the surviving text, which is otherwise near-complete including most supplementary tables.
- Figures (Figure 1, S-1, S-2) could not be analyzed directly; their content was inferred from captions and corresponding tables.
- The full fine-tuning control and selector-ablation sections rely on supplementary tables partially truncated in extraction; per-seed values cited were legible but a few table rows were cut.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.