Yifeng Ding, Hung Le, Songyang Han, Kangrui Ruan, Zhenghui Jin, Varun Kumar, Zijian Wang, Anoop Deoras. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

A popular approach to building agents using Language Models (LMs) involves iteratively prompting the LM, reflecting on its outputs, and updating the input prompts until the desired task is achieved. However, our analysis reveals two key shortcomings in the existing methods: $(i)$ limited exploration of the decision space due to repetitive reflections, which result in redundant inputs, and $(ii)$ an inability to leverage insights from previously solved tasks. To address these issues, we introduce DoT (Diversity of Thoughts), a novel framework that a) explicitly reduces redundant reflections to enhance decision-space exploration, and b) incorporates a task-agnostic memory component to enable knowledge retrieval from previously solved tasks—unlike current approaches that operate in isolation for each task. Through extensive experiments on a suite of programming benchmarks (HumanEval, MBPP, and LeetCodeHardGym) using a variety of LMs, DoT demonstrates up to a $\textbf{10}$% improvement in Pass@1 while maintaining cost-effectiveness. Furthermore, DoT is modular by design. For instance, when the diverse reflection module of DoT is integrated with existing methods like Tree of Thoughts (ToT), we observe a significant $\textbf{13}$% improvement on Game of 24 (one of the main benchmarks of ToT), highlighting the broad applicability and impact of our contributions across various reasoning tasks.

Sachit Kuhar, Wasi Uddin Ahmad, Zijian Wang, Nihal Jain, Haifeng Qian, Baishakhi Ray, Murali Krishna Ramanathan, Xiaofei Ma, Anoop Deoras. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Programming assistants powered by large language models have transformed software development, yet most benchmarks focus narrowly on code generation tasks. Recent efforts like InfiBench and StackEval attempt to address this gap using Stack Overflow data but remain limited to single-turn interactions in isolated contexts, require significant manual curation, and fail to represent complete project environments. We introduce CodeAssistBench (CAB), the first benchmark framework for evaluating multi-turn programming assistance in realistic settings that address questions grounded in actual codebases. Unlike existing programming Q&A benchmarks, CAB automatically generates scalable datasets from GitHub issues tagged with questions using configurable parameters (e.g., repository creation date, star count, programming languages), and includes automatic containerization of codebases for evaluation. It then evaluates models through simulated users in these containerized environments with full codebase access. Using this framework, we constructed a test set of 3,286 real-world programming questions across 214 repositories, spanning seven programming languages and diverse problem domains. Our evaluation of leading LLMs reveals a substantial capability gap: while models perform well on Stack Overflow questions with success rates of 70-83%, they resolve only up to 16.49% of CAB's issues from recent repositories (post-training cutoff). This discrepancy highlights the challenges of providing assistance in complex, project-specific contexts versus answering standalone questions. Our fully automated framework enables continuous benchmark expansion and is available at https://github.com/amazon-science/CodeAssistBench/.

It is common to reject undesired outputs of Large Language Models (LLMs); however, current methods to do so require an excessive amount of computation to re-sample after a rejection, or distort the distribution of outputs by constraining the output to highly improbable tokens. We present a method, Approximately Aligned Decoding (AprAD), to balance the distortion of the output distribution with computational efficiency, inspired by algorithms from the speculative decoding literature. AprAD allows for the generation of long sequences of text with difficult-to-satisfy constraints, while amplifying low probability outputs much less compared to existing methods. We show through a series of experiments that the task-specific performance of AprAD is comparable to methods that do not distort the output distribution, while being much more computationally efficient.

Software updates, including bug repair and feature additions, are frequent in modern applications but they often leave test suites outdated, resulting in undetected bugs and increased chances of system failures. A recent study by Meta revealed that 14%-22% of software failures stem from outdated tests that fail to reflect changes in the codebase. This highlights the need to keep tests in sync with code changes to ensure software reliability. In this paper, we present UTFix , a novel approach for repairing unit tests when their corresponding focal methods undergo changes. UTFix addresses two critical issues: assertion failure and reduced code coverage caused by changes in the focal method. Our approach leverages language models to repair unit tests by providing contextual information such as static code slices, dynamic code slices, and failure messages. We evaluate UTFix on our generated synthetic benchmark (Syn-Bench), and real-world benchmark. In our experiment, UTFix successfully repaired 89.2% of assertion failures and achieved 100% code coverage for 96 tests out of 369 unit tests. On the real-world benchmarks, UTFix repaired 60% of assertion failures while achieving 100% code coverage for 19 out of 30 unit tests. To the best of our knowledge, this is the first comprehensive study focused on unit test in evolving Python projects. Our contributions include the development of UTFix , the creation of Syn-Bench and real-world benchmarks, and the demonstration of the effectiveness of LLM-based methods in addressing unit test failures due to software evolution.

Large Language Models (LLMs) can exhibit considerable variation in the quality of their sampled outputs.Reranking and selecting the best generation from the sampled set is a popular way of obtaining strong gains in generation quality.In this paper, we present a novel approach for reranking LLM generations.Unlike other techniques that might involve additional inferences or training a specialized reranker, our approach relies on easy to compute pairwise statistics between the generations that have minimal compute overhead.We show that our approach can be formalized as an extension of self-consistency and analyze its performance in that framework, theoretically as well as via simulations.We show strong improvements for selecting the best k generations for code generation tasks as well as robust improvements for the best generation for the tasks of autoformalization, summarization, and translation.While our approach only assumes black-box access to LLMs, we show that additional access to token probabilities can improve performance even further.

In large language model training, input documents are typically concatenated together and then split into sequences of equal length to avoid padding tokens. Despite its efficiency, the concatenation approach compromises data integrity—it inevitably breaks many documents into incomplete pieces, leading to excessive truncations that hinder the model from learning to compose logically coherent and factually consistent content that is grounded on the complete context. To address the issue, we propose Best-fit Packing, a scalable and efficient method that packs documents into training sequences through length-aware combinatorial optimization. Our method completely eliminates unnecessary truncations while retaining the same training efficiency as concatenation. Empirical results from both text and code pre-training show that our method achieves superior performance (e.g., +4.7% on reading comprehension; +16.8% in context following; and +9.2% on program synthesis), and reduces closed-domain hallucination effectively by up to 58.3%.

We propose a new method to measure the task-specific accuracy of Retrieval-Augmented Large Language Models (RAG). Evaluation is performed by scoring the RAG on an automatically-generated synthetic exam composed of multiple choice questions based on the corpus of documents associated with the task. Our method is an automated, cost-efficient, interpretable, and robust strategy to select the optimal components for a RAG system. We leverage Item Response Theory (IRT) to estimate the quality of an exam and its informativeness on task-specific accuracy. IRT also provides a natural way to iteratively improve the exam by eliminating the exam questions that are not sufficiently informative about a model’s ability. We demonstrate our approach on four new open-ended Question-Answering tasks based on Arxiv abstracts, StackExchange questions, AWS DevOps troubleshooting guides, and SEC filings. In addition, our experiments reveal more general insights into factors impacting RAG performance like size, retrieval mechanism, prompting and fine-tuning. Most notably, our findings show that choosing the right retrieval algorithms often leads to bigger performance gains than simply using a larger language model.

Large models training is plagued by the intense compute cost and limited hardware memory. A practical solution is low-precision representation but is troubled by loss in numerical accuracy and unstable training rendering the model less useful. We argue that low-precision floating points can perform well provided the error is properly compensated at the critical locations in the training process. We propose Collage which utilizes multi-component float representation in low-precision to accurately perform operations with numerical errors accounted. To understand the impact of imprecision to training, we propose a simple and novel metric which tracks the lost information during training as well as differentiates various precision strategies. Our method works with commonly used low-precision such as half-precision ($16$-bit floating points) and can be naturally extended to work with even lower precision such as $8$-bit. Experimental results show that pre-training using Collage removes the requirement of using $32$-bit floating-point copies of the model and attains similar/better training performance compared to $(16, 32)$-bit mixed-precision strategy, with up to $3.7\times$ speedup and $\sim 15\%$ to $23\%$ less memory usage in practice. The code is available at https://github.com/amazon-science/collage.

In the domain of code generation, self-debugging is crucial. It allows LLMs to refine their generated code based on execution feedback. This is particularly important because generating correct solutions in one attempt proves challenging for complex tasks. Prior works on self-debugging mostly focus on prompting methods by providing LLMs with few-shot examples, which work poorly on small open-sourced LLMs. In this work, we propose LeDex, a training framework that significantly improves the self-debugging capability of LLMs. Intuitively, we observe that a chain of explanations on the wrong code followed by code refinement helps LLMs better analyze the wrong code and do refinement. We thus propose an automated pipeline to collect a high-quality dataset for code explanation and refinement by generating a number of explanations and refinement trajectories from the LLM itself or a larger teacher model and filtering via execution verification. We perform supervised fine-tuning (SFT) and further reinforcement learning (RL) on both success and failure trajectories with a novel reward design considering code explanation and refinement quality. SFT improves the pass@1 by up to 15.92\% and pass@10 by 9.30\% over four benchmarks. RL training brings additional up to 3.54\% improvement on pass@1 and 2.55\% improvement on pass@10. The trained LLMs show iterative refinement ability and can keep refining code continuously. Lastly, our human evaluation shows that the LLMs trained with our framework generate more useful code explanations and help developers better understand bugs in source code.

Recent studies on pre-trained vision/language models have demonstrated the practical benefit of a new, promising solution-building paradigm in AI where models can be pre-trained on broad data describing a generic task space and then adapted successfully to solve a wide range of downstream tasks, even when training data is severely limited (e.g., in zero- or few-shot learning scenarios). Inspired by such progress, we investigate in this paper the possibilities and challenges of adapting such a paradigm to the context of recommender systems, which is less investigated from the perspective of pre-trained model. In particular, we propose to develop a generic recommender that captures universal interaction patterns by training on generic user-item interaction data extracted from different domains, which can then be fast adapted to improve few-shot learning performance in unseen new domains (with limited data).

The unique capabilities of Large Language Models (LLMs), such as the natural language text generation ability, position them as strong candidates for providing explanation for recommendations. However, despite the size of the LLM, most existing models struggle to produce zero-shot explanations reliably. To address this issue, we propose a framework called Logic-Scaffolding, that combines the ideas of aspect-based explanation and chain-of-thought prompting to generate explanations through intermediate reasoning steps. In this paper, we share our experience in building the framework and present an interactive demonstration for exploring our results.

Modern recommender systems usually include separate recommendation carousels such as ‘trending now’ to list trending items and further boost their popularity, thereby attracting active users. Though widely useful, such ‘trending now’ carousels typically generate item lists based on simple heuristics, e.g., the number of interactions within a time interval, and therefore still leave much room for improvement. This paper aims to systematically study this under-explored but important problem from the new perspective of time series forecasting. We first provide a set of rigorous definitions related to item trendiness and formulate the trend recommendation task as a one-step time series forecasting problem. We then propose a deep latent variable model, dubbed Trend Recommender (TrendRec), to forecast items’ future trends and generate trending item lists. Furthermore, we design associated evaluation protocols for trend recommendation. Experiments on real-world datasets from various domains show that our TrendRec significantly outperforms the baselines, verifying our model’s effectiveness.

Item-to-Item (I2I) recommendation is an important function that suggests replacement or complement options for an item based on their functional similarities or synergies. To capture such item relationships effectively, the recommenders need to understand why subsets of items are co-viewed or co-purchased by the customers. Graph-based models, such as graph neural networks (GNNs), provide a natural framework to combine, ingest and extract valuable insights from such high-order item relationships. However, learning GNNs effectively for I2I requires ingesting a large amount of relational data, which might not always be available, especially in new, emerging market segments. To mitigate this data bottleneck, we postulate that recommendation patterns learned from existing market segments (with private data) could be adapted to build effective warm-start models for emerging ones. To achieve this, we introduce a personalized graph adaptation model based on GNNs to summarize, assemble and adapt recommendation patterns across market segments with heterogeneous customer behaviors into effective local models.

We study the problem of best-arm identification (BAI) in the fixed-budget setting with heterogeneous reward variances. We propose two variance-adaptive BAI algorithms for this setting: SHVar for known reward variances and SHAdaVar for unknown reward variances. Our algorithms rely on non-uniform budget allocations among the arms where the arms with higher reward variances are pulled more often than those with lower variances. The main algorithmic novelty is in the design of SHAdaVar, which allocates budget greedily based on overestimating the unknown reward variances. We bound probabilities of misidentifying the best arms in both SHVar and SHAdaVar. Our analyses rely on novel lower bounds on the number of pulls of an arm that do not require closed-form solutions to the budget allocation problem. Since one of our budget allocation problems is analogous to the optimal experiment design with unknown variances, we believe that our results are of a broad interest. Our experiments validate our theory, and show that SHVar and SHAdaVar outperform algorithms from prior works with analytical guarantees.

This paper studies the item-to-item recommendation problem in recommender systems from a new perspective of metric learning via implicit feedback. We develop and investigate a personalizable deep metric model that captures both the internal contents of items and how they were interacted with by users. There are two key challenges in learning such model. First, there is no explicit similarity annotation, which deviates from the assumption of most metric learning methods. Second, these approaches do not account for the fact that items are often represented by multiple sources of meta data and different users use different combinations of these sources to form their own notion of similarity. To address these challenges, we develop a new metric representation embedded as kernel parameters of a probabilistic model. This helps express the correlation between items that a user has interacted with, which can be used to predict user interaction with new items. Our approach hinges on the intuition that similar items induce similar interactions from the same user, thus fitting a metric-parameterized model to predict an implicit feedback signal could indirectly guide it towards finding the most suitable metric for each user. To this end, we also analyze how and when the proposed method is effective from a theoretical lens. Its empirical effectiveness is also demonstrated on several real-world datasets.

There are many studies that show using class-based features improves the performance of natural language processing (NLP) tasks such as syntactic part-of-speech tagging, dependency parsing, sentiment analysis, and slot filling in natural language understanding (NLU), but not much has been reported on the underlying reasons for the performance improvements. In this paper, we investigate the effects of the word class-based features for the exponential family of models specifically focusing on NLU tasks, and demonstrate that the performance improvements could be attributed to the regularization effect of the class-based features on the underlying model. Our hypothesis is based on empirical observation that shrinking the sum of parameter magnitudes in an exponential model tends to improve performance. We show on several semantic tagging tasks that there is a positive correlation between the model size reduction by the addition of the class-based features and the model performance on a held-out dataset. We also demonstrate that class-based features extracted from different data sources using alternate word clustering methods can individually contribute to the performance gain. Since the proposed features are generated in an unsupervised manner without significant computational overhead, the improvements in performance largely come for free and we show that such features provide gains for a wide range of tasks from semantic classification and slot tagging in NLU to named entity recognition (NER).

Models for statistical spoken language understanding (SLU) systems are conventionally trained using supervised discriminative training methods. In many cases, however, labeled data necessary for these supervised techniques is not readily available necessitating a laborious data collection and annotation effort. This often results into data sets that are not expansive enough to cover adequately all patterns of natural language phrases that occur in the target applications. Word embedding features alleviate data and feature sparsity issues by learning mathematical representation of words and word associations in the continuous space. In this work, we present techniques to obtain task and domain specific word embeddings and show their usefulness over those obtained from generic unsupervised data. We also show how we transfer these embeddings from one language to another enabling training of a multilingual spoken language understanding system.

In this paper we propose a set of class-based features that are generated in an unsupservised fashion to improve slot tagging with Conditional Random Fields (CRFs). The feature generation is based on the idea behind shrinkage based language models, where shrinking the sum of parameter magnitudes in an exponential model tends to improve performance. We use these features with CRFs and show that they consistently improve the slot tagging performance against baselines on several natural language understanding tasks. Since the proposed features are generated in an unsupervised manner without significant computational overhead, the improvements in performance comes for free and we expect that the same features may result in gains in other tagging tasks.

Conversational understanding systems, especially virtual personal assistants (VPAs), perform “targeted” natural language understanding, assuming their users stay within the walled gardens of covered domains, and back-off to generic web search otherwise. However, users usually do not know the concept of domains and sometimes simply do not distinguish the system from simple voice search. Hence it becomes an important problem to identify these rejected out-of-domain utterances which are actually intended for the VPA. This paper presents a study tackling this new task, showing that how one utters a request is more important for this task than what is uttered, resembling addressee detection or dialog act tagging. To this end, syntactic and semantic parse “structure” features are extracted in addition to lexical features to train a binary SVM classifier using a large number of random web search queries and VPA utterances from multiple domains. We present controlled experiments leaving one domain out and check the precision of the model when combined with unseen queries. Our results indicate that such structured features result in higher precision especially when the test domain bears little resemblance to the existing domains.

Applications of Deep Belief Nets (DBN) to various problems have been the subject of a number of recent studies ranging from image classification and speech recognition to audio classification. In this study we apply DBNs to a natural language understanding problem. The recent surge of activity in this area was largely spurred by the development of a greedy layer-wise pretraining method that uses an efficient learning algorithm called Contrastive Divergence (CD). CD allows DBNs to learn a multi-layer generative model from unlabeled data and the features discovered by this model are then used to initialize a feed-forward neural network which is fine-tuned with backpropagation. We compare a DBN-initialized neural network to three widely used text classification algorithms: Support Vector Machines (SVM), boosting and Maximum Entropy (MaxEnt). The plain DBN-based model gives a call-routing classification accuracy that is equal to the best of the other models. However, using additional unlabeled data for DBN pre-training and combining DBN-based learned features with the original features provides significant gains over SVMs, which, in turn, performed better than both MaxEnt and Boosting.

This paper investigates the use of deep belief networks (DBN) for semantic tagging, a sequence classification task, in spoken language understanding (SLU). We evaluate the performance of the DBN based sequence tagger on the well-studied ATIS task and compare our technique to conditional random fields (CRF), a state-of-the-art classifier for sequence classification. In conjunction with lexical and named entity features, we also use dependency parser based syntactic features and part of speech (POS) tags [1]. Under both noisy conditions (output of automatic speech recognition system) and clean conditions (manual transcriptions), our deep belief network based sequence tagger outperforms the best CRF based system described in [1] by an absolute 2% and 1% F-measure, respectively.Upon carrying out an analysis of cases where CRF and DBN models made different predictions, we observed that when discrete features are projected onto a continuous space during neural network training, the model learns to cluster these features leading to its improved generalization capability, relative to a CRF model, especially in cases where some features are either missing or noisy.

A challenge in large vocabulary spoken language understand-ing (SLU) is robustness to automatic speech recognition (ASR) errors. The state of the art approaches for semantic parsing rely on using discriminative sequence classification methods, such as conditional random fields (CRFs). Most dialog systems em-ploy a cascaded approach where the best hypotheses from the ASR system are fed into the following SLU system. In our pre-vious work, we have proposed the use of lattices towards joint recognition and parsing. In this paper, extending this idea, we propose to exploit word confusion networks (WCNs), compiled from ASR lattices for both CRF modeling and decoding. WCNs provide a compact representation of multiple aligned ASR hy-potheses, without compromising recognition accuracy. For slot filling, we show significant semantic parsing performance im-provements using WCNs compared to ASR 1-best output, ap-proximating the oracle path performance. Index Terms: conditional random field, semantic parsing, word confusion network, natural language understanding

Most Spoken Language Understanding (SLU) systems today employ a cascade approach, where the best hypothesis from Automatic Speech Recognizer (ASR) is fed into understanding modules such as slot sequence classifiers and intent detectors. The output of these modules is then further fed into downstream components such as interpreter and/or knowledge broker. These statistical models are usually trained individually to optimize the error rate of their respective output. In such approaches, errors from one module irreversibly propagates into other modules causing a serious degradation in the overall performance of the SLU system. Thus it is desirable to jointly optimize all the statistical models together. As a first step towards this, in this paper, we propose a joint decoding framework in which we predict the optimal word as well as slot sequence (semantic tag sequence) jointly given the input acoustic stream. Furthermore, the improved recognition output is then used for an utterance classification task, specifically, we focus on intent detection task. On a SLU task, we show 1.5% absolute reduction (7.6% relative reduction) in word error rate (WER) and 1.2% absolute improvement in F measure for slot prediction when compared to a very strong cascade baseline comprising of state-of-the-art large vocabulary ASR followed by conditional random field (CRF) based slot sequence tagger. Similarly, for intent detection, we show 1.2% absolute reduction (12% relative reduction) in classification error rate.

Most conversational understanding (CU) systems today employ a cascade approach, where the best hypothesis from automatic speech recognizer (ASR) is fed into spoken language understanding (SLU) module, whose best hypothesis is then fed into other systems such as interpreter or dialog manager. In such approaches, errors from one statistical module irreversibly propagates into another module causing a serious degradation in the overall performance of the conversational understanding system. Thus it is desirable to jointly optimize all the statistical modules together. As a first step towards this, in this paper, we propose a joint decoding framework in which we predict the optimal word as well as slot (semantic tag) sequence jointly given the input acoustic stream. On Microsoft’s CU system, we show 1.3 % absolute reduction in word error rate (WER) and 1.2% absolute improvement in F measure for slot prediction when compared to a very strong cascade baseline comprising of the state-of-the-art recognizer followed by a slot sequence tagger.

We describe how to effectively train neural network based language models on large data sets. Fast convergence during training and better overall performance is observed when the training data are sorted by their relevance. We introduce hash-based implementation of a maximum entropy model, that can be trained as a part of the neural network model. This leads to significant reduction of computational complexity. We achieved around 10% relative reduction of word error rate on English Broadcast News speech recognition task, against large 4-gram model trained on 400M tokens.

Long-span language models that capture syntax and semantics are seldom used in the first pass of large vocabulary continuous speech recognition systems due to the prohibitive search-space of sentence-hypotheses. Instead, an N-best list of hypotheses is created using tractable n-gram models, and rescored using the long-span models. It is shown in this paper that computationally tractable variational approximations of the long-span models are a better choice than standard n-gram models for first pass decoding. They not only result in a better first pass output, but also produce a lattice with a lower oracle word error rate, and rescoring the N-best list from such lattices with the long-span models requires a smaller N to attain the same accuracy. Empirical results on the WSJ, MIT Lectures, NIST 2007 Meeting Recognition and NIST 2001 Conversational Telephone Recognition data sets are presented to support these claims.

We present results obtained with several advanced language modeling techniques, including class based model, cache model, maximum entropy model, structured language model, random forest language model and several types of neural network based language models. We show results obtained after combining all these models by using linear interpolation. We conclude that for both small and moderately sized tasks, we obtain new state of the art results with combination of models, that is significantly better than performance of any individual model. Obtained perplexity reductions against Good-Turing trigram baseline are over 50% and against modified Kneser-Ney smoothed 5-gram over 40%. Index Terms: language modeling, neural networks, model combination, speech recognition

In this paper we investigate random forest based language model adaptation. Large amounts of out-of-domain data are used to grow the decision trees while very small amounts of in-domain data are used to prune them back, so that the structure of the trees are suitable for the desired domain while the probabilities in the tree nodes are reliably estimated. Extensive experiments are carried out and results are reported on a particular task of adapting Broadcast News language model to the MIT computer science lecture domain. We show 0.80% and 0.60% absolute WER improvement over language model interpolation and count merging techniques, respectively.

In this paper, we explore the model combination problem for rescoring Automatic Speech Recognition (ASR) hypotheses. We use minimum Empirical Bayes Risk for the optimization criterion and Deterministic Annealing techniques to search through the non-convex parameter space. Our experiments on the DARPA WSJ task using several different language models showed that our approach consistently outperforms the standard methods of model combination that optimize using 1-best hypothesis error.

Recently there has been a lot of interest in confusion network re-scoring using sophisticated and complex knowledge sources. Traditionally, re-scoring has been carried out by the N-best list method or by the lattices or confusion network dynamic programming method. Although the dynamic programming method is optimal, it allows for the incorporation of only Markov knowledge sources. N-best lists, on the other hand, can incorporate sentence level knowledge sources, but with increasing N, the re-scoring becomes computationally very intensive. In this paper, we present an elegant framework for confusion network re-scoring called `iterative decoding'. In it, integration of multiple and complex knowledge sources is not only easier but it also allows for much faster re-scoring as compared to the N-best list method. Experiments with language model re-scoring show that for comparable performance (in terms of word error rate (WER)) of iterative decoding and N-best list re-scoring, the search effort required by our method is 22 times less than that of the N-best list method.

This paper presents novel methods that integrate lexical prediction of non-verbalized punctuations with Viterbi decoding for Large Vocabulary Conversational Speech Recognition (LVCSR) in a single pass. We describe two different approaches- one based on a modified finite state machine representation of language models and one based on an extension of an LVCSR decoder. We discuss advantages over traditional punctuation prediction approaches based on post-processing of recognition hypotheses, including experimental evaluation of the proposed approach using a state-of-the-art LVCSR decoder. Experiments were performed on a medical documentation corpus and results demonstrate that the proposed methods yield improved punctuation prediction accuracy while at the same time reducing system complexity and memory requirements.