| Challenge: | Language models estimate a probability distribution over strings in a natural language . many recent linguistic studies have been incorrectly computing word probabilities . |
| Approach: | They propose to use the correct method to compute word probabilities . they highlight issues when relying on models that use end-of-word tokenisers . |
| Outcome: | Empirically, correcting the widespread bug affects measured outcomes in sentences and lexical optimisation analyses. |
Similar Papers
What Are the Odds? Language Models Are Capable of Probabilistic Reasoning (2024.emnlp-main)
Copied to clipboard
Akshay Paruchuri, Jake Garrison, Shun Liao, John Hernandez, Jacob Sunshine, Tim Althoff, Xin Liu, Daniel McDuff
| Challenge: | Language models (LMs) are capable of remarkably complex linguistic tasks, but numerical reasoning is an area in which they struggle. |
| Approach: | They evaluate the probabilistic reasoning capabilities of language models using idealized and real-world statistical distributions. |
| Outcome: | The proposed model can make inferences about distributions, even if assumptions are incorrect or misspecified. |
Clozing the Gap: Exploring Why Language Model Surprisal Outperforms Cloze Surprisal (2026.acl-long)
Copied to clipboard
| Challenge: | cloze task is being replaced with LM probabilities for how predictable a word is in its context . clozen task is used to measure how predictable word is compared to unexpected words . |
| Approach: | They propose to use LM probabilities to quantify how predictable a word is . they propose to test whether LMs are better at distinguishing semantically similar words . |
| Outcome: | The results show that LM probabilities outperform cloze probabilities for the right reasons . they also show that human-like prediction is more sensitive to LM probability distinctions . |
Leading Whitespaces of Language Models’ Subword Vocabulary Pose a Confound for Calculating Word Probabilities (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for aggregating word-by-word conditional probabilities into word probabilities are flawed . tokens in subword vocabulary of most language models have leading whitespaces and therefore do not define stop probabilities of words. |
| Approach: | They propose a method to reaccount the probability of trailing whitespace into that of the current word . they show lower estimates of garden-path effects in transitive/intransitive sentences . |
| Outcome: | The proposed method corrects the confound in word-by-word probabilities from LMs . the proposed method lowers garden-path effects in transitive/intransitive sentences . |
Language Model Probabilities are Not Calibrated in Numeric Contexts (2025.acl-long)
Copied to clipboard
Charles Lovering, Michael Krumdick, Viet Dac Lai, Varshini Reddy, Seth Ebner, Nilesh Kumar, Rik Koncel-Kedziorski, Chris Tanner
| Challenge: | Using language model outputs, we find that even in simple settings, the best LMs (1) are poorly calibrated and (2) have systematic biases. |
| Approach: | They argue that language model outputs should capture natural distributions over multiple options within their textual contexts. |
| Outcome: | The proposed model outputs are calibrated to the numeric content of their contexts. |
What Can String Probability Tell Us About Grammaticality? (2026.tacl-1)
Copied to clipboard
| Challenge: | linguistic theories have argued that language models have largely achieved grammatical competence, but they will assign non-zero probability to all strings. |
| Approach: | They propose a theoretical framework for analyzing string probabilities in linguistics based on simple assumptions about the generative process of corpus data. |
| Outcome: | The proposed framework makes three predictions using 280K sentence pairs in English and Chinese. |
Predict the Next Word: <Humans exhibit uncertainty in this task and language models _____> (2024.eacl-short)
Copied to clipboard
| Challenge: | Language models (LMs) are statistical models trained to assign probability to human-generated text. |
| Approach: | They evaluate language models' ability to reproduce variability that humans exhibit in the ‘next word prediction’ task. |
| Outcome: | The language models are trained to assign probability to human-generated text . they exhibit low calibration to human uncertainty, and advise against it . |
Subword models struggle with word learning, but surprisal hides it (2025.acl-short)
Copied to clipboard
| Challenge: | Subword LMs struggle to discern words and non-words with high accuracy, character LM models do this easily and consistently. |
| Approach: | They propose to model word learning in subword and character language models with the psycholinguistic lexical decision task. |
| Outcome: | The results suggest that word learning and syntactic learning are separable in character LMs. |
Distribution Prompting: Understanding the Expressivity of Language Models Through the Next-Token Distributions They Can Produce (2025.emnlp-main)
Copied to clipboard
| Challenge: | Autoregressive neural language models (LMs) generate a probability distribution over tokens at each time step given a prompt. |
| Approach: | They propose to find a prompt that induces LMs to output a distribution as close as possible to the target, using either soft or hard gradient-based prompt tuning. |
| Outcome: | The proposed model is able to generate a distribution as close as possible to a target given a prompt, and it can be used to approximate distributions with low or high entropy. |
A Survey of Confidence Estimation and Calibration in Large Language Models (2024.naacl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) have demonstrated impressive capabilities across a wide range of tasks in various domains, but they can be unreliable due to factual errors in their generations. |
| Approach: | They summarize recent advances in LLM confidence estimation and calibration and outline their main lessons learned. |
| Outcome: | The proposed methods can be used to assess the reliability of models and to calibrate them across tasks. |
Erratum: Measuring and Improving Consistency in Pretrained Language Models (2021.tacl-1)
Copied to clipboard
Yanai Elazar, Nora Kassner, Shauli Ravfogel, Abhilasha Ravichander, Eduard Hovy, Hinrich Schütze, Yoav Goldberg
| Challenge: | During production of this paper, an error was introduced to the formula on the bottom of the right column of page 1020. |
| Approach: | the formula was changed in the last two terms of the paper . |
| Outcome: | the correct formula is now available on the web. |