Disposition Distillation at Small Scale: A Three-Arc Negative Result

• URL: http://arxiv.org/abs/2604.11867v1
• Date: Mon, 13 Apr 2026 17:40:31 GMT
• Title: Disposition Distillation at Small Scale: A Three-Arc Negative Result
• Abstract summary: 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.
• Score: 0.0
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: We set out to train behavioral dispositions (self-verification, uncertainty acknowledgment, feedback integration) into small language models (0.6B to 2.3B effective parameters) through a four-stage all-MIT distillation pipeline, with follow-on experiments on inference-time attention-head interventions and a frozen-base confidence-gated sidecar. An 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. The HumanEval delta was a truncation artifact (n_predict=512) that inverted to -8.0 points at n_predict=1024; the MCAS gain disappeared under apples-to-apples scoring. That falsification triggered three subsequent arcs. Across (1) SFT/DPO LoRA on three model families and two domains, (2) inference-time attention-head tempering on o_proj, and (3) a training-free frozen-base sidecar reading the final-token hidden state h_last, we find no operator that moves judge-measured disposition without damaging content or collapsing into stylistic mimicry. The failure is consistent across five models (Qwen3-0.6B, Qwen3-1.7B, Qwen3.5-0.8B, Gemma 4 E2B, and SmolLM2-1.7B-Instruct). A within-distribution cross-validation pass (AUC=0.683) collapsed to chance on fresh prompts (AUC=0.516). 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. As an independent finding, Gemma 4 E2B exhibits near-complete confidence-correctness decoupling on the Chef domain (assertion asymmetry -0.009; the model asserts at 91% regardless of correctness).

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)
• When Do Internal Probes Beat Reading the Answer? Miscalibrated Readouts and Behavior-Concealed Knowledge in Language Models [0.0]
  A language model, asked to verify 1,200 logical conclusions, answers YES every time.<n>linear probes on its hidden states read the correct verdict at 0.96 AUC.<n>We ask where the verdict is lost, and find the dominant failure is a single scalar.<n>The diagnosis generalizes: across 90 semantic-label configurations of a five-model, three-family factorial, behavioral accuracy collapses onto a single function of threshold offset.
  arXiv  Detail & Related papers  (2026-09-04T00:26:45Z)
• Below the Noise Floor: Bimodal Seed Collapse and Distinct Failure Modes in Small-Model Knowledge Distillation [0.3525984766628294]
  On a 740-instance healthcare API routing task with a 1.5B Qwen student and a 20B teacher, we compare eight KD variants against supervised cross-entropy.<n>We find: (i) per-seed standard deviation ranges from 2.8 to 48.7 percentage points, swallowing every claimed KD gain below five points.
  arXiv  Detail & Related papers  (2026-08-27T21:40:19Z)
• Accurate Ensembles, Fragile Narratives: Multi-Scale Stacking and a Fidelity Audit of LLM-Generated Explanations for Credit Risk [0.0]
  Credit scoring increasingly relies on models whose decision logic cannot be read off their parameters.<n>A common proposal closes that gap with a language model: compute feature attributions, hand them to an LLM, and let it write the rationale.<n>We build such a system end to end and test whether the second half of the promise holds.
  arXiv  Detail & Related papers  (2026-08-08T13:22:14Z)
• Abliteration Is Not a Scalpel: Off-Target Effects of Refusal Removal on Decision Disposition Across Model Families [51.56484100374058]
  Abliteration - deleting a model's refusal direction from its weights - is the standard recipe behind popular "uncensored" open-weight models.<n>As a disposition probe we use 21,600 decisions under uncertainty, replayed through a frozen pipeline so the decision-layer model is the only variable.<n>Three effects replicate across both families (weeks-clustered CIs excluding zero)<n>A fourth effect reverses sign: the same operation makes Gemma-abliterated less confident and Qwen-abliterated more (family CIs non-overlapping)
  arXiv  Detail & Related papers  (2026-07-19T22:27:00Z)
• Silent Failures in Quantized LLM Reasoning: A Taxonomy-Based Analysis of Hollow Convergence and Failure Mode Shifts [0.0]
  We show that post-training quantization can silently alter how large language models reason even when task accuracy is preserved.<n>We classify 30,000 chain-of-thought outputs from five instruction-tuned LLMs across three quantization precisions.<n>We find that while accuracy is robust across precisions, Hollow Convergence shows a significant size-dependent shift under NF4.
  arXiv  Detail & Related papers  (2026-07-10T21:55:05Z)
• Right Makes Might: Aligning Verified Hidden States Empowers RL Reasoning [55.264863369127774]
  Current methods reduce each correct rollout to a single reward bit, ignoring the geometric structure shared among their hidden states.<n>We propose Hidden-Align, an auxiliary loss function that aligns the last-layer hidden states of correct rollouts at the anchor token during RL training, with zero overhead in both training and inference.<n>On eight mathematical reasoning benchmarks, Hidden-Align improves average pass@1 over the DAPO baseline by 3.8, 6.2, and 5.4 percentage points on Qwen3-1.7B, 4B, and 14B respectively, with consistent pass@k gains across all three scales.
  arXiv  Detail & Related papers  (2026-06-02T06:51:15Z)
• 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)
• Distributional Energy-Based Models for Uncertainty-Aware Structured LLM Reasoning [40.342912574072024]
  Large Language Models produce structured outputs such as travel plans or code solutions.<n>Individual reasoning steps may appear correct while the output as a whole violates budgets, fails test cases, or contradicts earlier deductions.<n>We propose a learned quality scorer with deterministic analytical constraint penalties for verifying structured LLM outputs.
  arXiv  Detail & Related papers  (2026-05-15T17:08:27Z)
• The Extrapolation Cliff in On-Policy Distillation of Near-Deterministic Structured Outputs [52.709361620508595]
  ListOPD brings a student to in-domain with an 8B-SFT baseline at one-fifth the parameters.<n>On Amazon Fashion, three pre-registered tests--a fine-grid cliff interval test, and a small-clip cross-prediction--fall within their locked prediction windows--with the small-clip value matching the closed-form prediction below grid resolution.
  arXiv  Detail & Related papers  (2026-05-09T06:48:00Z)
• Distilling Self-Consistency into Verbal Confidence: A Pre-Registered Negative Result and Post-Hoc Rescue on Gemma 3 4B [0.0]
  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.
  arXiv  Detail & Related papers  (2026-04-27T05:53:26Z)
• 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)
• The Geometry of Harmful Intent: Training-Free Anomaly Detection via Angular Deviation in LLM Residual Streams [0.0]
  We present LatentBiopsy, a training-free method for detecting harmful prompts by analysing the geometry of residual-stream activations in large language models.<n>We evaluate two complete model triplets from the Qwen3.5-0.8B and Qwen2.5-0.5B families.<n>Across all six variants, LatentBiopsy achieves AUROC $geq$0.937 for harmful-vs-normative detection and AUROC = 1.000 for discriminating harmful from benign-aggressive prompts.
  arXiv  Detail & Related papers  (2026-03-28T21:19:58Z)
• Self-Anchoring Calibration Drift in Large Language Models: How Multi-Turn Conversations Reshape Model Confidence [0.0]
  Self-Anchoring Condition Drift (SACD) is a hypothesized tendency for large language models to show systematic changes in expressed confidence when building iteratively on their own prior outputs across multi-turn conversations.<n>We compare Claude Sonnet 4.6, Gemini 3.1 Pro, and GPT-5.2 across 150 questions spanning factual, technical, and open-ended domains, using three conditions: single-turn baseline (A), multi-turn self-anchoring (B), and independent repetition control (C)<n>Results reveal a complex, model-heterogeneous pattern that partially diverges from pre-registered hypotheses.
  arXiv  Detail & Related papers  (2026-03-01T19:27:52Z)
• Information Fidelity in Tool-Using LLM Agents: A Martingale Analysis of the Model Context Protocol [69.11739400975445]
  We introduce the first theoretical framework for analyzing error accumulation in Model Context Protocol (MCP) agents.<n>We show that cumulative distortion exhibits linear growth and high-probability deviations bounded by $O(sqrtT)$.<n>Key findings include: semantic weighting reduces distortion by 80%, and periodic re-grounding approximately every 9 steps suffices for error control.
  arXiv  Detail & Related papers  (2026-02-10T21:08:53Z)
• Prompt Injection Evaluations: Refusal Boundary Instability and Artifact-Dependent Compliance in GPT-4-Series Models [0.0]
  We evaluate two models, GPT-4.1 and GPT-4o, using 3,274 perturbation runs derived from refusal-inducing prompt injection attempts.<n>We find that artifact type is a stronger predictor of refusal failure than perturbation style.<n>We conclude that refusal behavior is a probabilistic, artifact-dependent boundary phenomenon rather than a stable binary property.
  arXiv  Detail & Related papers  (2026-01-25T17:14:33Z)
• 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)
• Detecting AI Hallucinations in Finance: An Information-Theoretic Method Cuts Hallucination Rate by 92% [4.693270291878929]
  Large language models (LLMs) produce fluent but unsupported answers - hallucinations.<n>We propose ECLIPSE, a framework that treats hallucination as a mismatch between a model's semantic entropy and the capacity of available evidence.
  arXiv  Detail & Related papers  (2025-12-02T05:25:48Z)
• RefineBench: Evaluating Refinement Capability of Language Models via Checklists [71.02281792867531]
  We evaluate two refinement modes: guided refinement and self-refinement.<n>In guided refinement, both proprietary LMs and large open-weight LMs can leverage targeted feedback to refine responses to near-perfect levels within five turns.<n>These findings suggest that frontier LMs require breakthroughs to self-refine their incorrect responses.
  arXiv  Detail & Related papers  (2025-11-27T07:20:52Z)
• SEAL: Steerable Reasoning Calibration of Large Language Models for Free [58.931194824519935]
  Large Language Models (LLMs) have demonstrated compelling capabilities for complex reasoning tasks via the extended chain-of-thought (CoT) reasoning mechanism.<n>Recent studies reveal substantial redundancy in the CoT reasoning traces, which negatively impacts model performance.<n>We introduce SEAL, a training-free approach that seamlessly calibrates the CoT process, improving accuracy while demonstrating significant efficiency gains.
  arXiv  Detail & Related papers  (2025-04-07T02:42:07Z)

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.