Distilling Self-Consistency into Verbal Confidence: A Pre-Registered Negative Result and Post-Hoc Rescue on Gemma 3 4B

• URL: http://arxiv.org/abs/2604.24070v1
• Date: Mon, 27 Apr 2026 05:53:26 GMT
• Title: Distilling Self-Consistency into Verbal Confidence: A Pre-Registered Negative Result and Post-Hoc Rescue on Gemma 3 4B
• Abstract summary: Small instruct-tuned LLMs produce degenerate verbal confidence under minimal elicitation.<n>We test whether confidence-conditioned supervised fine-tuning with self-consistency-derived targets can close the gap between internal information and verbal readout.
• Score: 0.0
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Small instruct-tuned LLMs produce degenerate verbal confidence under minimal elicitation: ceiling rates above 95%, near-chance Type-2 AUROC, and Invalid validity profiles. We test whether confidence-conditioned supervised fine-tuning (CSFT) with self-consistency-derived targets can close the gap between internal information and verbal readout. A pre-registered Phase 0 protocol on Gemma 3 4B-it with a modal filter restricting training to items with correct modal answers produced a negative result: AUROC2 dropped from 0.554 to 0.509 due to label-entropy collapse in the training targets. An exploratory rescue removed the filter, training on all 2,000 calibration items. This produced a binary verbal correctness discriminator with AUROC2 = 0.774 on held-out TriviaQA, compressing a 10-sample self-consistency signal (AUROC2 = 0.999) into a single-pass readout exceeding logit entropy (0.701). The shuffled-target control showed no improvement (0.501). On MMLU, accuracy improved from 54.2% to 77.4% with the shuffled model at baseline (56.1%), supporting a target-dependent interpretation. The result is exploratory, binary rather than continuously calibrated, and observed at a single scale. It identifies two design lessons: confidence training requires label entropy, and correct targets regularise output format.

Related papers

• ReDraft, Don't Just Distill: Reference-Driven Revision for Continual VLLM Post-Training [66.72708893922605]
  Post-training of large multimodal models should add new capabilities while preserving those from pre-training.<n>SFT gives explicit target supervision that learns a task from near-zero accuracy, but its off-policy targets move the model far enough to cause forgetting.<n>We introduce ReDraft, which obtains both from the model's own failures.
  arXiv  Detail & Related papers  (2026-09-15T04:59:03Z)
• The Model Proposes, the Code Disposes: A Pre-Registered Ablation of a Verifier-and-Acceptance Stage in an LLM-Orchestrated Offensive-Security Agent [0.0]
  We evaluate whether a verifier-and-acceptance stage - a model verifier whose verdicts are enforced by deterministic code - changes what an LLM-driven offensive-security agent reports.<n>We report a 15-run exploratory pilot, a pre-registered 20-run confirmatory ablation, and a pre-registered 2 x 2 factorial study with 40 runs across two deliberately vulnerable lab targets.<n>The results support a narrow conclusion: the verifier changes what the system ships, while deterministic code supplies enforcement and auditability.
  arXiv  Detail & Related papers  (2026-09-14T17:13:05Z)
• Empirical Evaluation of Membership Inference Attacks on NLP Text Classifiers: A Baseline Study on SST-2 [0.0]
  Membership inference attacks (MIAs) try to determine whether a specific record was used to train a model.<n>This paper presents a benchmark of membership inference vulnerability for text classification on the GLUE SST-2 sentiment dataset.
  arXiv  Detail & Related papers  (2026-09-10T00:53:20Z)
• Support Operation Factorization: Compositional Readout of Frozen Vision Encoders under Controlled Interventions [87.95757610455173]
  We introduce an injectively aligned leave-one-cell-out protocol over support x operation grids and SO-OPF.<n>We separate two questions that aggregate scores conflate: whether the carrier composes held-out bindings when the grid is known, and whether that grid can be recovered from flat cell labels.
  arXiv  Detail & Related papers  (2026-08-06T15:38:55Z)
• SVR: Self-Verifying Refinement via Joint Verdict-Confidence Reinforcement Learning for Adaptive Test-Time Compute [62.3458279176813]
  Self-Verifying Refinement is an oracle-free multi-turn reinforcement learning framework.<n>It learns to use self-verification as a compute-control policy.<n>It achieves a macro-average accuracy of 0.563 with only 2.99 inference turns on average.
  arXiv  Detail & Related papers  (2026-07-30T16:20:58Z)
• Amplifying, Not Learning: Fine-Tuned AI Text Detectors Amplify a Pretrained Direction [51.56484100374058]
  Text detectors amplify a pretrained typicality axis; they do not construct an AI-vs-human boundary.<n>On raw encoders before any task supervision, projecting onto centroid(AI)-centroid(HC3) achieves NYT-vs-HC3 AUROC 0.806/0.944/0.834 across three architectures.<n>On RoBERTa-base, raw projection exceeds fine-tuning; on RoBERTa-base, full fine-tuning reduces discrimination below raw on both fluent-formal populations tested.
  arXiv  Detail & Related papers  (2026-05-20T19:08:38Z)
• Verbal Confidence Saturation in 3-9B Open-Weight Instruction-Tuned LLMs: A Pre-Registered Psychometric Validity Screen [0.0]
  We tested whether seven instruction-tuned open-weight models produce verbalised confidence that meets minimal validity criteria.<n>We found Categorical elicitation did not rescue validity.<n> Token-level logprobability did not usefully predict verbalised confidence under the observed variance regime.
  arXiv  Detail & Related papers  (2026-04-24T04:45:21Z)
• The Illusion of Certainty: Decoupling Capability and Calibration in On-Policy Distillation [67.26315138466312]
  We propose a calibration-aware OPD framework, CaOPD, that estimates empirical confidence from model rollouts.<n>Our findings highlight that capability distillation does not imply confidence, and that confidence should be treated as an essential objective in post-training.
  arXiv  Detail & Related papers  (2026-04-18T04:43:40Z)
• The Verification Tax: Fundamental Limits of AI Auditing in the Rare-Error Regime [0.0]
  Most cited calibration result in deep learning -- post-temperature-scaling ECE of 0.012 on CIFAR-100 -- is below the statistical noise floor.<n>We prove this is not a failure of the experiment but a law: the minimax rate for estimating calibration error with model error rate epsilon is Theta((Lepsilon/m)2/3), and no estimator can beat it.
  arXiv  Detail & Related papers  (2026-04-14T16:48:24Z)
• Calibrated Confidence Estimation for Tabular Question Answering [0.0]
  All models are severely overconfident (smooth ECE 0.35-0.64 versus 0.10-0.15 reported for textual QA)<n>A consistent self-evaluation versus perturbation dichotomy replicates across both benchmarks and all four fully-covered models.<n>A secondary contribution, structure-aware recalibration, improves AUROC by percentage points over standard post-hoc methods.
  arXiv  Detail & Related papers  (2026-04-14T09:16:53Z)
• Disposition Distillation at Small Scale: A Three-Arc Negative Result [0.0]
  Internal draft reported +33.9-point MCAS and +15.3-point HumanEval gains on a Qwen3-0.6B student; a second-pass sanity check falsified both numbers before publication.<n>We find no operator that moves judge-measured disposition without damaging content or collapsing into stylistic mimicry.<n>We contribute a three-arc negative result with mechanism, a two-failure-mode taxonomy for linear h_last probes, and an honest falsification pipeline that converts the class of false positives we ourselves produced into publishable negatives.
  arXiv  Detail & Related papers  (2026-04-13T17:40:31Z)
• CoRefine: Confidence-Guided Self-Refinement for Adaptive Test-Time Compute [10.548368675645403]
  CoRefine is a confidence-guided self-refinement method that achieves competitive accuracy using a fraction of the tokens.<n>The controller consumes full-trace confidence to decide whether to halt, re-examine, or try a different approach.<n>We extend this to CoRefine-Tree, a hybrid sequential-parallel variant that adaptively balances exploration and exploitation.
  arXiv  Detail & Related papers  (2026-02-09T17:44:41Z)
• CARE What Fails: Contrastive Anchored-REflection for Verifiable Multimodal [84.71254539482369]
  Group-relative reinforcement learning with verifiable rewards (RLVR) often wastes the most informative data it already has the failures.<n>We present CARE, a failure-centric post-training framework for multimodal reasoning that turns errors into supervision.<n> CARE improves accuracy and training smoothness while explicitly increasing the share of learning signal that comes from failures.
  arXiv  Detail & Related papers  (2025-12-22T16:34:21Z)
• Annotation-Efficient Universal Honesty Alignment [70.05453324928955]
  Existing methods either rely on training-free confidence estimation or training-based calibration with correctness annotations.<n>Elicitation-Then-Calibration (EliCal) is a two-stage framework that first elicits internal confidence using inexpensive self-consistency supervision, then calibrates this confidence with a small set of correctness annotations.<n>EliCal achieves near-optimal alignment with only 1k correctness annotations (0.18% of full supervision) and better alignment performance on unseen MMLU tasks than the calibration-only baseline.
  arXiv  Detail & Related papers  (2025-10-20T13:05:22Z)
• Robust Fine-tuning of Zero-shot Models via Variance Reduction [56.360865951192324]
  When fine-tuning zero-shot models, our desideratum is for the fine-tuned model to excel in both in-distribution (ID) and out-of-distribution (OOD) We propose a sample-wise ensembling technique that can simultaneously attain the best ID and OOD accuracy without the trade-offs.
  arXiv  Detail & Related papers  (2024-11-11T13:13:39Z)
• Adversarial Training with Rectified Rejection [114.83821848791206]
  We propose to use true confidence (T-Con) as a certainty oracle, and learn to predict T-Con by rectifying confidence. We prove that under mild conditions, a rectified confidence (R-Con) rejector and a confidence rejector can be coupled to distinguish any wrongly classified input from correctly classified ones.
  arXiv  Detail & Related papers  (2021-05-31T08:24:53Z)
• Semi-supervised Contrastive Learning with Similarity Co-calibration [72.38187308270135]
  We propose a novel training strategy, termed as Semi-supervised Contrastive Learning (SsCL) SsCL combines the well-known contrastive loss in self-supervised learning with the cross entropy loss in semi-supervised learning. We show that SsCL produces more discriminative representation and is beneficial to few shot learning.
  arXiv  Detail & Related papers  (2021-05-16T09:13:56Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.