It is shown that post-training quantization can silently alter how large language models reason even when task accuracy is preserved, and Hollow Convergence is established as a deployment-relevant failure mode that standard evaluation pipelines cannot catch.
Abstract
We show that post-training quantization can silently alter how large language models reason even when task accuracy is preserved. Using a six-category failure taxonomy validated by two independent human annotators (Cohen's $\kappa$ = 0.906), we classify 30,000 chain-of-thought outputs from five instruction-tuned LLMs (3B--14B parameters) across three quantization precisions (FP32, FP16, NF4) and four reasoning benchmarks. We find that while accuracy is robust across precisions (maximum 3.1 pp drop), Hollow Convergence (correct answers reached through incomplete or unverifiable reasoning) shows a significant size-dependent shift under NF4, dropping sharply for the two smallest models tested but remaining invariant for models at 12B parameters and above. This effect is also benchmark-specific: GSM8K is categorically immune while LogiQA and ARC-Challenge show the largest shifts. Furthermore, under NF4, Shortcut Collapse rises from 44% to 78% of wrong-answer failures in LLaMA 3.2-3B while Confidence Snowballing collapses from 15.8% to near zero, a qualitative shift invisible to accuracy metrics. Finally, we show Hollow Convergence cannot be reliably detected from surface-level text features (best F1 = 0.53), establishing it as a deployment-relevant failure mode that standard evaluation pipelines cannot catch.
A graded, multi-family, failure-aware framework for stress-testing reasoning models that exposes structure that an aggregate score hides and exposes consistent weaknesses across all models.
Accuracy changes after language-model self-revision are usually interpreted as changes in reasoning. We show this can fail at the answer-extraction boundary, and test the failure causally rather than only observationally. Across Qwen3.5 (0.8B-9B), Gemma-4-12B, and two frontier models via API (Tencent Hy3, Nvidia Nemotron-3-Ultra-550B) in 29 primary cells plus a frontier arm, we decompose the always-revise accuracy shift into a content margin (both answers parseable) and format-recovery/loss margins (parseability changes). On 12 cells with meaningful unparseable-answer rates, format effects exceed content effects (Wilcoxon p=1.7e-3). To test this causally, we force already-generated reasoning through grammar-constrained decoding so every answer is parseable by construction: across 14 cells this closes a median 71% of the gap between the naive total effect and the content-margin estimate, with two cells converging exactly and a residual on the two largest-effect cells reported rather than dismissed. A clustered model confirms floor-scale (0.8B/2B) models have far higher odds of content-level change and harm than capable-scale models (p<1e-7). Replicating a cited confidence-gating protocol verbatim on Qwen3.5 does not reproduce its reported gain and shows the same near-zero content margin. A frontier check on much larger models shows format-dominance intensifying with scale: content margin is exactly zero in all 5 cells despite total effects up to +0.275, though this arm is lower-powered. The calibration-floor criterion on the content margin reveals a squeeze: floor-scale cells have headroom but insufficient signal, capable-scale cells have signal but little headroom; only one cell is marginally viable, with negligible sealed-holdout gain. Content is a minority share of what the field has measured as self-correction. We release the instrument, code, and derived results.
Large language models (LLMs) are increasingly used as graders, verifiers, and process auditors, but most mathematical evaluations still emphasize final-answer accuracy. This can obscure whether a model can verify a non-canonical but valid solution trace. We introduce a controlled linear-equation benchmark for evaluating LLMs in the evaluator role. Each instance asks the model to judge final-answer correctness, step-level trace correctness, and the first incorrect step. Our evaluation of state-of-the-art open LLMs reveals a significant robustness gap: models that accurately evaluate canonical solutions often fail when presented with perturbed but logically equivalent variants. Across GPT-OSS 20B, Qwen3-14B, and Phi-4-Reasoning, base models perform well on canonical traces but degrade substantially on perturbed traces, especially for error localization. On valid perturbed traces, base-model false-rejection rates reach 75.6-85.3%, showing strong sensitivity to canonical solution form. Supervised fine-tuning, distillation, and test-time compute improve robustness in some settings, but gains are model dependent and can trade off against canonical performance. The results show that reliable process-level verification remains challenging, and evaluator robustness should be measured separately from solver accuracy, even in a simple algebraic domain with exact ground truth.
A two-layer evaluation framework that separates scorer-independent execution evidence, including termination, answer exposure, parseability, and completion length, from scorer-dependent correctness is introduced, demonstrating that accuracy conflates execution case mix with verification policy.
Inference-time activation steering enables behavioral control of large language models without parameter modification, while post-training quantization reduces memory and compute costs for deployment. Despite their growing convergence in practice, the interaction between these two techniques remains uncharacterized. We systematically study activation steering under weight-only quantization (INT8 and NF4) across four open-weight 7-9B models and two behavioral targets: judged sentiment and judge-free reasoning length. Using an iso-effect framework that compares capability costs at matched behavioral effect, we find that sentiment steering survives quantization intact. After correcting a GSM8K parser artifact with a uniform v2.3.1 rescore, the pooled INT8 contrast is -0.010 (90% CI [-0.026, +0.007]), descriptively Equivalent under the preregistered three-label rule, while NF4 remains Inconclusive at -0.017 ([-0.067, +0.033]). In contrast, reasoning length exhibits a surprising asymmetric dose-response: lengthening is graded but terminates in cap-runaway and collapse, while shortening is a step function with only 12-30% shortening (model-dependent) before discontinuous failure. We expose a methodological pitfall: the naive iso-effect ladder anchors on the collapse floor for floor-bounded targets, and we introduce a censored construction that restores interpretable crossings. We also quantify a substantial baseline capability shift for Mistral-NF4 (0.545 to 0.365 GSM8K at alpha=0), demonstrating that compression can dominate the steering intervention. Despite this, steering vectors remain highly collinear with their FP16 siblings (cosine similarity 0.989-0.998 for INT8, 0.945-0.990 for NF4), confirming that the behavioral direction survives quantization even when the cost structure does not. All code and data are released.
A probe corpus of 42 retracted, fraudulent, and pseudoscientific papers is paired with a methodology for eliciting and scoring single-shot model engagement with each paper's framing, indicating an urgent need for guardrail infrastructure for scientific deployment of language models.
V. Rodionov, Shamil Assylbekov· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
