Does Reasoning Preserve Alignment? On the Trustworthiness of Large Reasoning Models

• URL: http://arxiv.org/abs/2606.11046v1
• Date: Tue, 09 Jun 2026 16:14:27 GMT
• Title: Does Reasoning Preserve Alignment? On the Trustworthiness of Large Reasoning Models
• Abstract summary: We compare reasoning models produced via supervised fine-tuning, RL-based post-training, and distillation against instruction-tuned baselines.<n>We observe that reasoning models often improve on reasoning benchmarks but exhibit alignment regressions.<n>These regressions are consistent with behavioral drift from the instruction-tuned baseline, measured by KL divergence.
• Score: 55.788110316999166
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Instruction-tuned LLMs are increasingly converted into reasoning models through post-training to improve multi-step task performance. This conversion is usually optimized for reasoning accuracy, without explicitly preserving the alignment behavior of the instruction-tuned model, such as safe refusal, bias avoidance, and privacy protection. We ask: does this conversion preserve alignment? We study this question through a trustworthiness audit and find that it is not behavior-preserving by default. For a systematic analysis, we compare reasoning models produced via supervised fine-tuning, RL-based post-training, and distillation against matched instruction-tuned baselines across six trustworthiness dimensions: safety, toxicity, stereotyping and bias, machine ethics, privacy, and out-of-distribution robustness. We observe that reasoning models often improve on reasoning benchmarks but exhibit alignment regressions, including increased toxicity, amplified stereotyping, miscalibrated refusal, and contextual privacy leakage. These regressions are consistent with behavioral drift from the instruction-tuned baseline, measured by KL divergence. Overall, our results point to the broader conclusion that trustworthiness metrics are essential for evaluating reasoning models and should be reported alongside gains in reasoning capability.

Related papers

• Invascal: Inverse-Vacuity Self-Calibration for Uncertainty-Aware LiDAR Range-View Semantic Segmentation [3.0338740183736106]
  We introduce a novel, architecture-agnostic uncertainty-aware Adapter Head.<n>It decomposes the prediction into a Preference Head for class ranking and a Strength Head that refines uncertainty assessment.<n>We propose our inverse-vacuity self-calibration objective (Invascal), which directly supervises the strength signal to produce reliable and well-calibrated uncertainty estimates.
  arXiv  Detail & Related papers  (2026-05-20T16:39:29Z)
• Confidence-Aware Alignment Makes Reasoning LLMs More Reliable [65.44962502963378]
  We introduce CASPO, a framework that aligns token-level confidence with step-wise logical correctness without training a separate reward model.<n>During inference, we propose Confidence-aware Thought (CaT), which dynamically prunes uncertain reasoning branches with negligible O(V) latency.<n>Experiments across ten benchmarks and multiple model families show that CASPO consistently improves reasoning reliability and inference efficiency.
  arXiv  Detail & Related papers  (2026-05-08T07:08:25Z)
• Calibration Drift Under Reasoning: How Chain-of-Thought Budgets Induce Overconfidence in Large Language Models [0.0]
  Chain-of-thought (CoT) reasoning is widely used to improve accuracy and reliability.<n>We show that increasing the reasoning budget beyond a task-specific threshold can cause models to become systematically overconfident.<n>We propose a Hypothesis Lock-In model based on autoregressive generation to explain this behavior.
  arXiv  Detail & Related papers  (2026-04-24T04:46:16Z)
• 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)
• Quantifying Self-Preservation Bias in Large Language Models [9.590157416396194]
  We introduce the emphTwo-role Benchmark for Self-Preservation.<n>It measures how often role identity overrides objective utility.<n>We observe that in low-improvement regimes, models exploit the interpretive slack to post-hoc rationalization.
  arXiv  Detail & Related papers  (2026-04-02T15:38:31Z)
• THINKSAFE: Self-Generated Safety Alignment for Reasoning Models [60.10077024249373]
  We propose ThinkSafe, a framework that restores safety alignment without external teachers.<n>Our key insight is that while compliance suppresses safety mechanisms, models often retain latent knowledge to identify harm.<n> Experiments on DeepSeek-R1-Distill and Qwen3 show ThinkSafe significantly improves safety while preserving reasoning proficiency.
  arXiv  Detail & Related papers  (2026-01-30T16:31:02Z)
• Large Reasoning Models Learn Better Alignment from Flawed Thinking [56.08883934423522]
  Large reasoning models (LRMs) "think" by generating structured chain-of-thought (CoT) before producing a final answer.<n>We propose RECAP, a principled reinforcement learning (RL) method for post-training that explicitly teaches models to override flawed reasoning trajectories.
  arXiv  Detail & Related papers  (2025-10-01T14:15:43Z)
• Is Reasoning All You Need? Probing Bias in the Age of Reasoning Language Models [0.0]
  Reasoning Language Models (RLMs) have gained traction for their ability to perform complex, multi-step reasoning tasks.<n>While these capabilities promise improved reliability, their impact on robustness to social biases remains unclear.<n>We leverage the CLEAR-Bias benchmark to investigate the adversarial robustness of RLMs to bias elicitation.
  arXiv  Detail & Related papers  (2025-07-03T17:01:53Z)
• Selective Learning: Towards Robust Calibration with Dynamic Regularization [79.92633587914659]
  Miscalibration in deep learning refers to there is a discrepancy between the predicted confidence and performance. We introduce Dynamic Regularization (DReg) which aims to learn what should be learned during training thereby circumventing the confidence adjusting trade-off.
  arXiv  Detail & Related papers  (2024-02-13T11:25:20Z)

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.