Trust Your Guide Only When Certain: Uncertainty-Aware Sparse Alignment at Inference Time

• URL: http://arxiv.org/abs/2609.00624v1
• Date: Tue, 01 Sep 2026 03:05:49 GMT
• Title: Trust Your Guide Only When Certain: Uncertainty-Aware Sparse Alignment at Inference Time
• Abstract summary: Weak supervisors exhibit pervasive high entropy across the vast majority of tokens.<n>Trust-based Uncertainty Sparse Alignment (TUSA) reframes alignment as a dynamic arbitration process.<n>Experiments show that TUSA consistently improves both safety alignment and general helpfulness.
• Score: 16.36057592850375
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: A prominent paradigm in inference-time alignment employs lightweight supervisors to steer Large Language Models (LLMs). Through empirical analysis, we identify a structural mismatch in this paradigm: weak supervisors exhibit pervasive high entropy across the vast majority of tokens, yet prevailing dense intervention approaches mandate supervision at every decoding step. This leads to frequent low-confidence interventions that can disrupt valid base-model reasoning and incur substantial utility costs. To resolve this, we propose TUSA (Trust-based Uncertainty Sparse Alignment). Moving away from continuous oversight, TUSA reframes alignment as a dynamic arbitration process, introducing an uncertainty-aware arbiter that authorizes intervention only when two conditions are met: the supervisor is confident and the token is semantically salient. This mechanism effectively filters out uncertainty-driven noise and redundant supervision. Extensive experiments across multiple models and benchmarks show that TUSA consistently improves both safety alignment and general helpfulness. By bypassing approximately 50% of alignment steps, it not only enhances safety preference by up to 15.6%, but also boosts general preference rates by up to 12.0% compared to the dense baseline, demonstrating that selective, high-precision alignment can outperform continuous supervision.

Related papers

• Does Reasoning Preserve Alignment? On the Trustworthiness of Large Reasoning Models [55.788110316999166]
  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.
  arXiv  Detail & Related papers  (2026-06-09T16:14:27Z)
• Unveiling the Entropy Dynamics of Chain-of-Thought Reasoning [81.57028614960576]
  We show that the Confidence Region possesses two critical properties: 1) High Reliability -- answers in the confidence region become highly accurate and stable, and 2) High Redundancy -- models generate unnecessary tokens long after reaching the correct answer.<n>These properties unlock more efficient and reliable inference strategies.
  arXiv  Detail & Related papers  (2026-06-01T10:11:14Z)
• 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)
• Uncertainty-Aware Reward Discounting for Mitigating Reward Hacking [0.0]
  Real-world objectives are often uncertain, context-dependent, and internally inconsistent.<n>This mismatch can lead to alignment failures such as reward hacking, over-optimization, and overconfident behavior.<n>We introduce a dual-source uncertainty-aware reward framework that explicitly models both uncertainty in value estimation and uncertainty in human preferences.
  arXiv  Detail & Related papers  (2026-04-29T07:14:01Z)
• Mitigating Error Amplification in Fast Adversarial Training [58.74042726356826]
  Fast Adversarial Training (FAT) has proven effective in enhancing model robustness by encouraging networks to learn perturbation-invariant representations.<n>FAT often suffers from catastrophic overfitting (CO), where the model overfits to the training attack and fails to generalize to unseen ones.<n>We propose a Distribution-aware Dynamic Guidance (DDG) strategy that dynamically adjusts both the perturbation budget and supervision signal.
  arXiv  Detail & Related papers  (2026-04-27T11:23:18Z)
• Less Approximates More: Harmonizing Performance and Confidence Faithfulness via Hybrid Post-Training for High-Stakes Tasks [52.81286869496811]
  Large language models are increasingly deployed in high-stakes tasks, where confident yet incorrect inferences may cause real-world harm.<n>We propose to jointly optimize unsupervised Reinforcement Learning from Internal Feedback with reasoning-trace-guided Reasoning Distillation.<n>We introduce HyTuning, a hybrid post-training framework that adaptively reweights RD and RLIF via a PRG-style metric.
  arXiv  Detail & Related papers  (2026-04-09T16:50:11Z)
• Agentic Uncertainty Quantification [76.94013626702183]
  We propose a unified Dual-Process Agentic UQ (AUQ) framework that transforms verbalized uncertainty into active, bi-directional control signals.<n>Our architecture comprises two complementary mechanisms: System 1 (Uncertainty-Aware Memory, UAM), which implicitly propagates verbalized confidence and semantic explanations to prevent blind decision-making; and System 2 (Uncertainty-Aware Reflection, UAR), which utilizes these explanations as rational cues to trigger targeted inference-time resolution only when necessary.
  arXiv  Detail & Related papers  (2026-01-22T07:16:26Z)
• Decision-Aware Trust Signal Alignment for SOC Alert Triage [0.0]
  The present paper presents a decision-sensitive trust signal correspondence scheme of SOC alert triage.<n>The framework combines confidence that has been calibrated, lightweight uncertainty cues, and cost-sensitive decision thresholds into coherent decision-support layer.<n>We show that false negatives are greatly amplified by the presence of misaligned displays of confidence, whereas cost weighted loss decreases by orders of magnitude between models with decision aligned trust signals.
  arXiv  Detail & Related papers  (2026-01-08T01:41:54Z)
• Core Safety Values for Provably Corrigible Agents [2.6451153531057985]
  We introduce the first implementable framework for corrigibility, with provable guarantees in multi-step, partially observed environments.<n>Our framework replaces a single reward with five *structurally separate* utility heads.<n>For open-ended settings where adversaries can modify the agent, we prove that deciding whether an arbitrary post-hack agent will ever violate corrigibility is undecidable.
  arXiv  Detail & Related papers  (2025-07-28T16:19:25Z)

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.