Frontier LLMs are updated frequently and typically outperform their predecessors in aggregate. But aggregate gains say little about individual samples: an update can still cause sample-level regression, where a response correct under the old model becomes incorrect under the new one. This paper studies how to predict such regressions from signals available at inference time. We compare single-model signals (confidence, logit margin, attention entropy) against cross-version signals (output KL divergence, likelihood drift, token-level KL, representation drift) under a unified added-value test that isolates each signal's gain over a confidence baseline. Across six benchmarks in three task families (multiple-choice question answering, or MCQ; math reasoning; code generation) and six model update pairs, we find that (1) signal effectiveness is task-dependent: confidence is strongest on MCQ and simpler math, while likelihood/KL signals give the most frequent gains on harder math and code; (2) no signal is universally best across model updates either; and (3) some cross-version signals stay informative even when confidence fails, including without labels, which supports a proof-of-concept selective fallback that routes high-risk samples back to the old model. Practitioners can use these task-level patterns to choose which regression signal to trust for a given update. Code is available at https://github.com/jiashengsally/llm-regression-signals.
High-confidence errors in large language models are often treated as evidence of fragile internal inference. We study a different possibility: stable miscalibration, where a confident wrong answer remains locally stable under small perturbations. We combine two diagnostics: a label-aware output-level audit score that ranks domains by confidence variation and overconfident mistakes under a forced-answer baseline, and an internal sensitivity probe that measures hidden-state movement. On a multi-domain binary factual audit set, this audit score tracks where abstention-aware self-critique reduces decision loss, although direct labeled baselines rank the same gain more strongly. Internally, self-critical prompting consistently reduces hidden-state sensitivity across layers in three open-weight models. This supports prompt-induced local stabilization rather than a purely output-level abstention pattern, but it does not imply calibration: audit-defined overconfident errors are not clearly more locally sensitive than confidently correct answers, so some high-confidence errors may be stable and miscalibrated rather than simply fragile.
We investigate whether language models internally track the value of their current trajectory, defined as the likelihood that their ongoing strategy will achieve their goals. Using synthetic, in-context reinforcement learning data, we construct a "value" axis for Qwen3-8B. We find that activations along this axis distinguish between high vs. low verbalized confidence, rollouts without and with backtracking, and correct vs. corrupted code. Steering towards high value causally suppresses self-correction and reduces explanatory verbosity, while steering towards low value induces backtracking and exploration. We demonstrate that direct preference optimization (DPO) can increase the internal value of rewarded behaviors (e.g. use a certain word), causing the model to act more confidently after exhibiting them. Finally, we apply the value axis to study in-the-wild settings. For example, we find that Qwen assigns low value to politically sensitive chat queries after post-training and that supervised fine-tuning increases internal confidence within the training domain. Our results suggest that language models linearly encode an estimate of expected goal success that modulates their confidence in pursuing a direction.
Modern agent systems can turn uncertainty into overconfidence. Fragile upstream decisions are often exposed to downstream components as clean intermediate artifacts, while the uncertainty behind those decisions is lost at the interface. As a result, local ambiguity can become system-level error amplification. We argue that this reveals an interface bottleneck in agent uncertainty propagation: uncertainty does not propagate simply because a trajectory contains uncertain steps; it propagates only when it survives the handoff between components. We define uncertain decision handoff as the transfer of an intermediate decision made under uncertainty, and identify confidence laundering as a failure mode in which fragile upstream states are repackaged as procedurally valid artifacts that downstream agents over-trust. To address this bottleneck, we propose latent uncertainty as an uncertainty-bearing carrier attached to decision handoffs. Rather than replacing text with hidden states, latent uncertainty aims to preserve pre-commitment fragility in a form that downstream components can use. This position shifts agent uncertainty propagation from step-wise uncertainty estimation toward uncertainty-preserving interface design for more recoverable agent systems.
Machine learning is increasingly used in critical domains, where both predictions and their associated confidence levels influence important decisions. To enhance transparency in such scenarios, it is important to understand why a model is confident or uncertain about its predictions. Recent logic-based approaches provide abductive explanations, minimal subsets of features sufficient to preserve the predicted class, with correctness guarantees. However, these methods focus solely on classification behavior and may produce explanations that cover instances with low predictive confidence. In this work, we introduce the concept of Minimum Confidence Threshold (MCT), which quantifies the weakest confidence guarantee provided by an abductive explanation. Building upon this concept, we propose confidence-aware abductive explanations, which preserve not only the predicted class but also a user-specified confidence guarantee. We formulate MCT computation as an optimization problem and introduce an algorithm for generating minimal explanations that satisfy a desired confidence threshold. We evaluate the proposed framework on boosted trees for binary classification, although the approach is applicable to other machine learning models that provide confidence scores. Experimental results show that traditional abductive explanations often provide substantially weaker confidence guarantees than the confidence associated with the explained instance itself. In contrast, confidence-aware explanations consistently improve the minimum confidence guaranteed by an explanation while requiring only a modest increase in explanation length. These properties make the proposed approach particularly suitable for applications where both predictive correctness and confidence are essential for trustworthy decision making.