Charles Lovering, Michael Krumdick, Viet Dac Lai, Varshini Reddy, Seth Ebner, Nilesh Kumar, Rik Koncel-Kedziorski, Chris Tanner. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Reward models are central to aligning LLMs with human preferences, but they are costly to train, requiring large-scale human-labeled preference data and powerful pretrained LLM backbones. Meanwhile, the increasing availability of high-quality synthetic instruction-following datasets raises the question: can simpler, reference-based metrics serve as viable alternatives to reward models during RL-based alignment? In this paper, we show first that BLEU, a basic string-matching metric, surprisingly matches strong reward models in agreement with human preferences on general instruction-following datasets. Based on this insight, we develop BLEUBERI, a method that first identifies challenging instructions and then applies Group Relative Policy Optimization (GRPO) using BLEU directly as the reward function. We demonstrate that BLEUBERI-trained models are competitive with models trained via reward model-guided RL across four challenging instruction-following benchmarks and three different base language models. A human evaluation further supports that the quality of BLEUBERI model outputs is on par with those from reward model-aligned models. Moreover, BLEUBERI models generate outputs that are more factually grounded than competing methods. Overall, we show that given access to high-quality reference outputs (easily obtained via existing instruction-following datasets or synthetic data generation), string matching-based metrics are cheap yet effective proxies for reward models during alignment. We release our code and data at https://github.com/lilakk/BLEUBERI.

Michael Krumdick, Rik Koncel-Kedziorski, Viet Dac Lai, Varshini Reddy, Charles Lovering, Chris Tanner. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Tokenization is a foundational step in natural language processing (NLP) tasks, bridging raw text and language models.Existing tokenization approaches like Byte-Pair Encoding (BPE) originate from the field of data compression, and it has been suggested that the effectiveness of BPE stems from its ability to condense text into a relatively small number of tokens.We test the hypothesis that fewer tokens lead to better downstream performance by introducing PathPiece, a new tokenizer that segments a document's text into the minimum number of tokens for a given vocabulary.Through extensive experimentation we find this hypothesis not to be the case, casting doubt on the understanding of the reasons for effective tokenization.To examine which other factors play a role, we evaluate design decisions across all three phases of tokenization: pre-tokenization, vocabulary construction, and segmentation, offering new insights into the design of effective tokenizers.Specifically, we illustrate the importance of pretokenization and the benefits of using BPE to initialize vocabulary construction.We train 64 language models with varying tokenization, ranging in size from 350M to 2.4B parameters, all of which are made publicly available.

Event coreference models cluster event mentions pertaining to the same real-world event.Recent models rely on contextualized representations to recognize coreference among lexically or contextually similar mentions.However, models typically fail to leverage commonsense inferences, which is particularly limiting for resolving lexically-divergent mentions.We propose a model that extends event mentions with temporal commonsense inferences.Given a complex sentence with multiple events, e.g., "The man killed his wife and got arrested", with the target event "arrested", our model generates plausible events that happen before the target event -such as "the police arrived", and after it, such as "he was sentenced".We show that incorporating such inferences into an existing event coreference model improves its performance, and we analyze the coreferences in which such temporal knowledge is required. ① Mention Detection② Baseline Pairwise Scorer (spent, recovering): 0.14 (spent, gunshots): 0.05 (spent, shot): 0.1 (gunshots, shot): 0.82 … gunshots, shot, … recovering hospitalized spent Document 2The third coworker, Bryant Dalton, 39, spent two weeks in the hospital and still is recovering from gunshots to the neck and shoulder, prosecutors said. Document 1Bryant Dalton, 39, was shot in the neck and is hospitalized in good condition.… Document N … (spent, recovering): 0.14 (spent, gunshots): 0.05 (spent, shot): 0.1 (gunshots, shot): 0.82 … ③ Agglomerative Clustering gunshots, shot, … recovering spent, hospitalized spent, recovering, gunshots, shot, hospitalized ② Our Pairwise Scorer (spent, hospitalized): 0.1 (spent, hospitalized): 0.75 f(ctx spent , ctx hospitalized ) f(ctx spent , ctx hospitalized , cs spent , cs hospitalized ,)

Sign Language Recognition (SLR) aims to improve the equity of communication with the hearing impaired. However, SLR typically relies on having recorded videos of the signer. We develop a more natural solution by fitting a signer with arm sensors and classifying the sensor signals directly into language. We refer to this task as Sensors-to-Sign-Language (STSL). While existing STSL systems demonstrate effectiveness with small vocabularies of fewer than 100 words, we aim to determine if STSL can scale to larger, more realistic lexicons. For this purpose, we introduce a new dataset, SignBank, which consists of exactly 6,000 signs, spans 558 distinct words from 15 different novice signers, and constitutes the largest such dataset. By using a simple but effective model for STSL, we demonstrate a strong baseline performance on SignBank. Notably, despite our model having trained on only four signings of each word, it is able to correctly classify new signings with 95.1% accuracy (out of 558 candidate words). This work enables and motivates further development of lightweight, wearable hardware and real-time modelling for SLR.

Text documents of varying nature (e.g., summary documents written by analysts or published, scientific papers) often cite others as a means of providing evidence to support a claim, attributing credit, or referring the reader to related work.We address the problem of predicting a document's cited sources by introducing a novel, discriminative approach which combines a content-based generative model (LDA) with author-based features.Further, our classifier is able to learn the importance and quality of each topic within our corpus -which can be useful beyond this task -and preliminary results suggest its metric is competitive with other standard metrics (Topic Coherence).Our flagship system, Logit-Expanded, provides state-of-the-art performance on the largest corpus ever used for this task.