How Large Language Models Get Stuck: Early structure with persistent errors

• URL: http://arxiv.org/abs/2603.00359v1
• Date: Fri, 27 Feb 2026 22:49:11 GMT
• Title: How Large Language Models Get Stuck: Early structure with persistent errors
• Authors: Alokesh Manna, William Snyder, Whitney Tabor,
• Abstract summary: We trained Meta's OPT model on the 100M word BabyLM dataset.<n>We evaluated it on the BLiMP benchmark, which consists of 67 classes.<n>We tested the model's preference for grammatical over ungrammatical sentences.
• Score: 0.0
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Linguistic insights may help make Large Language Model (LLM) training more efficient. We trained Meta's OPT model on the 100M word BabyLM dataset, and evaluated it on the BLiMP benchmark, which consists of 67 classes, each defined by sentence pairs that differ in a targeted syntactic or semantic rule violation. We tested the model's preference for grammatical over ungrammatical sentences across training iterations and grammatical types. In nearly one-third of the BLiMP classes, OPT fails to consistently assign a higher likelihood to grammatical sentences, even after extensive training. When it fails, it often establishes a clear (erroneous) separation of the likelihoods at an early stage of processing and sustains this to the end of our training phase. We hypothesize that this mis-categorization is costly because it creates entrenched biases that must, eventually, be reversed in order for the model to perform well. We probe this phenomenon using a mixture of qualitative (based on linguistic theory and the theory of Deep Learning) and quantitative (based on numerical testing) assessments. Our qualitative assessments indicate that only some BLiMP tests are meaningful guides. We conclude by articulating a hypothesis, the Bigram Hypothesis, which claims that the learning process will exhibit erroneous entrenchment if bigram statistics bias the model toward wrong distinctions early in training, and we describe a method (in progress) of testing the hypothesis on appropriately selected BLiMP classes.

Related papers

• LLM Reasoning Predicts When Models Are Right: Evidence from Coding Classroom Discourse [0.18268488712787334]
  Large Language Models (LLMs) are increasingly deployed to automatically label and analyze educational dialogue at scale.<n>We investigate whether reasoning generated by LLMs can be used to predict the correctness of a model's own predictions.<n>We analyze 30,300 teacher utterances from classroom dialogue, each labeled by multiple state-of-the-art LLMs with an instructional move construct and an accompanying reasoning.
  arXiv  Detail & Related papers  (2026-02-10T14:38:13Z)
• Mitigating Reversal Curse in Large Language Models via Semantic-aware Permutation Training [57.771940716189114]
  We show that large language models (LLMs) suffer from the "reversal curse" The root cause of the reversal curse lies in the different word order between the training and inference stage. We propose Semantic-aware Permutation Training (SPT) to address this issue.
  arXiv  Detail & Related papers  (2024-03-01T18:55:20Z)
• Causally Testing Gender Bias in LLMs: A Case Study on Occupational Bias [33.99768156365231]
  We introduce a causal formulation for bias measurement in generative language models.<n>We propose a benchmark called OccuGender, with a bias-measuring procedure to investigate occupational gender bias.<n>The results show that these models exhibit substantial occupational gender bias.
  arXiv  Detail & Related papers  (2022-12-20T22:41:24Z)
• Beyond Distributional Hypothesis: Let Language Models Learn Meaning-Text Correspondence [45.9949173746044]
  We show that large-size pre-trained language models (PLMs) do not satisfy the logical negation property (LNP) We propose a novel intermediate training task, names meaning-matching, designed to directly learn a meaning-text correspondence. We find that the task enables PLMs to learn lexical semantic information.
  arXiv  Detail & Related papers  (2022-05-08T08:37:36Z)
• Uncertainty Estimation for Language Reward Models [5.33024001730262]
  Language models can learn a range of capabilities from unsupervised training on text corpora. It is often easier for humans to choose between options than to provide labeled data, and prior work has achieved state-of-the-art performance by training a reward model from such preference comparisons. We seek to address these problems via uncertainty estimation, which can improve sample efficiency and robustness using active learning and risk-averse reinforcement learning.
  arXiv  Detail & Related papers  (2022-03-14T20:13:21Z)
• Self-Normalized Importance Sampling for Neural Language Modeling [97.96857871187052]
  In this work, we propose self-normalized importance sampling. Compared to our previous work, the criteria considered in this work are self-normalized and there is no need to further conduct a correction step. We show that our proposed self-normalized importance sampling is competitive in both research-oriented and production-oriented automatic speech recognition tasks.
  arXiv  Detail & Related papers  (2021-11-11T16:57:53Z)
• NoiER: An Approach for Training more Reliable Fine-TunedDownstream Task Models [54.184609286094044]
  We propose noise entropy regularisation (NoiER) as an efficient learning paradigm that solves the problem without auxiliary models and additional data. The proposed approach improved traditional OOD detection evaluation metrics by 55% on average compared to the original fine-tuned models.
  arXiv  Detail & Related papers  (2021-08-29T06:58:28Z)
• On Sampling-Based Training Criteria for Neural Language Modeling [97.35284042981675]
  We consider Monte Carlo sampling, importance sampling, a novel method we call compensated partial summation, and noise contrastive estimation. We show that all these sampling methods can perform equally well, as long as we correct for the intended class posterior probabilities. Experimental results in language modeling and automatic speech recognition on Switchboard and LibriSpeech support our claim.
  arXiv  Detail & Related papers  (2021-04-21T12:55:52Z)
• Masked Language Modeling and the Distributional Hypothesis: Order Word Matters Pre-training for Little [74.49773960145681]
  A possible explanation for the impressive performance of masked language model (MLM)-training is that such models have learned to represent the syntactic structures prevalent in NLP pipelines. In this paper, we propose a different explanation: pre-trains succeed on downstream tasks almost entirely due to their ability to model higher-order word co-occurrence statistics. Our results show that purely distributional information largely explains the success of pre-training, and underscore the importance of curating challenging evaluation datasets that require deeper linguistic knowledge.
  arXiv  Detail & Related papers  (2021-04-14T06:30:36Z)
• Pre-training Is (Almost) All You Need: An Application to Commonsense Reasoning [61.32992639292889]
  Fine-tuning of pre-trained transformer models has become the standard approach for solving common NLP tasks. We introduce a new scoring method that casts a plausibility ranking task in a full-text format. We show that our method provides a much more stable training phase across random restarts.
  arXiv  Detail & Related papers  (2020-04-29T10:54:40Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.