Named Entity Recognition (NER) serves as a foundational component in many natural language processing (NLP) pipelines. However, current NER models typically output a single predicted label sequence without any accompanying measure of uncertainty, leaving downstream applications vulnerable to cascading errors. In this paper, we introduce a general framework for adapting sequence-labeling-based NER models to produce uncertainty-aware prediction sets. These prediction sets are collections of full-sentence labelings that are guaranteed to contain the correct labeling with a user-specified confidence level. This approach serves a role analogous to confidence intervals in classical statistics by providing formal guarantees about the reliability of model predictions. Our method builds on conformal prediction, which offers finite-sample coverage guarantees under minimal assumptions. We design efficient nonconformity scoring functions to construct efficient, well-calibrated prediction sets that support both unconditional and class-conditional coverage. This framework accounts for heterogeneity across sentence length, language, entity type, and number of entities within a sentence. Empirical experiments on four NER models across three benchmark datasets demonstrate the broad applicability, validity, and efficiency of the proposed methods.

It is widely recognized that reinforcement learning (RL) fine-tuning of large language models often leads to diversity collapse, where outputs lack variety. Prior work has proposed a range of heuristics to counteract this effect, but these methods are ad hoc: they frequently trade off correctness for diversity, their effectiveness varies across tasks, and in some cases they even contradict one another. In this work, we place these observations on a rigorous foundation. We first provide a formal proof of why RL fine-tuning exhibits diversity collapse via a selection and reinforcement bias. Next, we make a key observation that any reward modification to address diversity collapse only needs to be applied on the correct trajectories. Building directly on this analysis, we introduce a principled method -- differential smoothing -- that provably improves both correctness and diversity, outperforming vanilla RL as well as widely used entropy-based heuristics. Our theory precisely characterizes when existing heuristics help and why they fail, while showing that differential smoothing is universally superior. Extensive experiments with models from 1B to 7B parameters, across domains including CountDown and real-world mathematical reasoning, demonstrate consistent gains. Differential smoothing improves both Pass@1 and Pass@k, with up to 6.7% improvements on AIME24 dataset.

We study the problem of entropy calibration, which asks whether a language model's entropy over generations matches its log loss on human text. Past work found that models are miscalibrated, with entropy per step increasing (and text quality decreasing) as generations grow longer. This error accumulation is a fundamental problem in autoregressive models, and the standard solution is to truncate the distribution, which improves text quality at the cost of diversity. In this paper, we ask: is miscalibration likely to improve with scale, and is it theoretically possible to calibrate without tradeoffs? To build intuition, we first study a simplified theoretical setting to characterize the scaling behavior of miscalibration with respect to dataset size. We find that the scaling behavior depends on the power law exponent of the data distribution -- in particular, for a power law exponent close to 1, the scaling exponent is close to 0, meaning that miscalibration improves very slowly with scale. Next, we measure miscalibration empirically in language models ranging from 0.5B to 70B parameters. We find that the observed scaling behavior is similar to what is predicted by the simplified setting: our fitted scaling exponents for text are close to 0, meaning that larger models accumulate error at a similar rate as smaller ones. This scaling (or, lack thereof) provides one explanation for why we sample from larger models with similar amounts of truncation as smaller models, even though the larger models are of higher quality. However, truncation is not a satisfying solution because it comes at the cost of increased log loss. In theory, is it even possible to reduce entropy while preserving log loss? We prove that it is possible, if we assume access to a black box which can fit models to predict the future entropy of text.

There are many potential SSTs not discussed here.

This section provides the full proof of the theorems stated in the main paper. Specifically, for a "softened" sample We proceed to show the correctness of the backward pass. Therefore, the claim at the beginning of this paragraph holds. Plug in Eq. (14), we have context Finally, we note that Alg. 1 and 2 both run in time O ( |p| |D|). A.2 Useful Lemmas This section provides several useful lemmas that are later used in the proof of Thm. 4. Lemma 1. F ollowing this decomposition, we construct Alg. 4 that computes the entropy of every nodes in a Consider the example PC in Figure 1.

This section provides the full proof of the theorems stated in the main paper. Specifically, for a "softened" sample We proceed to show the correctness of the backward pass. Therefore, the claim at the beginning of this paragraph holds. Plug in Eq. (14), we have context Finally, we note that Alg. 1 and 2 both run in time O ( |p| |D|). A.2 Useful Lemmas This section provides several useful lemmas that are later used in the proof of Thm. 4. Lemma 1. F ollowing this decomposition, we construct Alg. 4 that computes the entropy of every nodes in a Consider the example PC in Figure 1.

D.2 Additional experiment results Comparison across related baselines.

Dragon fruit, renowned for its nutritional benefits and economic value, has experienced rising global demand due to its affordability and local availability. As dragon fruit cultivation expands, efficient pre- and post-harvest quality inspection has become essential for improving agricultural productivity and minimizing post-harvest losses. This study presents DragonFruitQualityNet, a lightweight Convolutional Neural Network (CNN) optimized for real-time quality assessment of dragon fruits on mobile devices. We curated a diverse dataset of 13,789 images, integrating self-collected samples with public datasets (dataset from Mendeley Data), and classified them into four categories: fresh, immature, mature, and defective fruits to ensure robust model training. The proposed model achieves an impressive 93.98% accuracy, outperforming existing methods in fruit quality classification. To facilitate practical adoption, we embedded the model into an intuitive mobile application, enabling farmers and agricultural stakeholders to conduct on-device, real-time quality inspections. This research provides an accurate, efficient, and scalable AI-driven solution for dragon fruit quality control, supporting digital agriculture and empowering smallholder farmers with accessible technology. By bridging the gap between research and real-world application, our work advances post-harvest management and promotes sustainable farming practices.

Episodic memory -- the ability to recall specific events grounded in time and space -- is a cornerstone of human cognition, enabling not only coherent storytelling, but also planning and decision-making. Despite their remarkable capabilities, Large Language Models (LLMs) lack a robust mechanism for episodic memory: we argue that integrating episodic memory capabilities into LLM is essential for advancing AI towards human-like cognition, increasing their potential to reason consistently and ground their output in real-world episodic events, hence avoiding confabulations. To address this challenge, we introduce a comprehensive framework to model and evaluate LLM episodic memory capabilities. Drawing inspiration from cognitive science, we develop a structured approach to represent episodic events, encapsulating temporal and spatial contexts, involved entities, and detailed descriptions. We synthesize a unique episodic memory benchmark, free from contamination, and release open source code and datasets to assess LLM performance across various recall and episodic reasoning tasks. Our evaluation of state-of-the-art models, including GPT-4 and Claude variants, Llama 3.1, and o1-mini, reveals that even the most advanced LLMs struggle with episodic memory tasks, particularly when dealing with multiple related events or complex spatio-temporal relationships -- even in contexts as short as 10k-100k tokens.