As the chain-of-thought reasoning capabilities of large language models improve, evaluating and calibrating their reasoning confidence is becoming increasingly important for quantifying the uncertaint...
The paper introduces Divergent Token Confidence (DTC), a method that estimates large language model confidence by counting tokens where two models strongly disagree during decoding. DTC uses Jensen-Shannon divergence between next-token distributions along the same reasoning trajectory and shows a near-negative correlation with answer accuracy. Experiments on multiple model families and six mathematical benchmarks demonstrate that DTC improves calibration over traditional probability-based and verbalized baselines, achieving lower expected calibration errors in both white-box and black-box settings.
The paper introduces two token-error supervision methods for sparse mixture-of-experts (MoE) large language models, aligning routing affinities with token-level cross-entropy loss. The first method predicts an error score per expert and uses it to adjust affinities before top‑K selection, while the second directly aligns router affinities to the model’s objective without extra heads or inference changes. Experiments on Granite and ARC‑Challenge show accuracy gains of about 2.3–2.94 percentage points over a parameter‑matched baseline, preserving the native sparse execution budget.
The paper investigates hallucination detection in black‑box large language models by leveraging two accessible signals: semantic entropy, which captures disagreement among sampled response meanings, and token‑level uncertainty derived from log‑probabilities. It introduces a TopK aggregation technique, a hybrid CoCoA method combining uncertainty with semantic dissimilarity, and two supervised approaches—Gated and Stacked—that integrate token and semantic features. Across seven benchmarks and four language models, the supervised Stacked method performs best in many cases, while TopK and CoCoA remain competitive without labeled data, though all methods require careful threshold calibration.
The paper investigates the "score granularity gap" in black-box large language model (LLM) classifiers, asking how finely a confidence score can be thresholded for deployment. By comparing seven confidence construction methods across 25 model-dataset pairs, the authors find that single-shot verbalized confidence, when properly converted to a probability, ranks well but offers only a few distinct threshold values, limiting operational flexibility. The study also shows that multi-query aggregation can improve weak models but may harm strong ones, and provides concrete guidance for deployment trade-offs.
arXiv:2603. 24929v2 Announce Type: replace Abstract: Understanding and quantifying uncertainty in large language model (LLM) outputs is critical for reliable deployment.
arXiv:2609.37594v1 Announce Type: new Abstract: Uncertainty estimates tell us how unsure a model is, but not why. Without knowing which parts of an input influences a model's uncertainty, we cannot t...
arXiv:2609.35831v1 Announce Type: new Abstract: Modern large language models now support context windows of more than one million tokens, which has raised the question of whether retrieval-augmented...
arXiv:2608. 19323v1 Announce Type: cross Abstract: Uncertainty quantification (UQ) is essential for the safe deployment of large language models (LLMs).
arXiv:2607. 25600v1 Announce Type: cross Abstract: Retrieval-augmented generation improves knowledge-intensive question answering, but indiscriminate retrieval can introduce irrelevant evidence and unnecessary computation.
The paper presents the first large‑scale benchmark for uncertainty quantification (UQ) calibration in long‑form scientific question answering, evaluating four UQ methods on 685,000 responses from up to 20 large language models across seven datasets. It shows that instruction tuning leads to token‑level probability polarization, undermining token‑level uncertainty signals, while reasoning model families differ in how they handle this effect. Only semantic consistency—consistency of the final answer—provides well‑calibrated outputs, demonstrating that semantic calibration remains robust in multi‑step, dependency‑rich reasoning.
arXiv:2609.22206v1 Announce Type: cross Abstract: Multimodal Large Language Models (MLLMs) have achieved remarkable performance across a wide range of multimodal tasks, yet understanding and quantify...