mlsanityData to 14 Oct 2026

Next-Token Prediction Should be Ambiguity-Sensitive: A Meta-Learning Perspective

arXiv 2506.16288 ↗First posted 19 Jun 2025cs.LG cs.AIcited by 0

Not on the map. Why?

Abstract (arXiv)

The rapid adaptation ability of auto-regressive foundation models is often attributed to the diversity of their pre-training data. This is because, from a Bayesian standpoint, minimizing prediction error in such settings requires integrating over all plausible latent hypotheses consistent with observations. While this behavior is desirable in principle, it often proves too ambitious in practice: under high ambiguity, the number of plausible latent alternatives makes Bayes-optimal prediction computationally intractable. Cognitive science has long recognized this limitation, suggesting that under such conditions, heuristics or information-seeking strategies are preferable to exhaustive inference. Translating this insight to next-token prediction, we hypothesize that low- and high-ambiguity predictions pose different computational demands, making ambiguity-agnostic next-token prediction a detrimental inductive bias. To test this, we introduce MetaHMM, a synthetic sequence meta-learning benchmark with rich compositional structure and a tractable Bayesian oracle. We show that Transformers indeed struggle with high-ambiguity predictions across model sizes. Motivated by cognitive theories, we propose a method to convert pre-trained models into Monte Carlo predictors that decouple task inference from token prediction. Preliminary results show substantial gains in ambiguous contexts through improved capacity allocation and test-time scalable inference, though challenges remain.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 41 of its 57 references that are in the map. First posted on arXiv 19 Jun 2025.

Builds on

41 of 57 references.

References · 41

  1. 2017Attention Is All You Needcited by 60,361
  2. 2020Language Models are Few-Shot Learnerscited by 32,419
  3. 2022Chain-of-Thought Prompting Elicits Reasoning in Large Language Modelscited by 12,459
  4. 2020Retrieval-Augmented Generation for Knowledge-Intensive NLP Taskscited by 9,297
  5. 2022ReAct: Synergizing Reasoning and Acting in Language Modelscited by 8,695
  6. 2018Representation Learning with Contrastive Predictive Codingcited by 7,777
  7. 2022Scalable Diffusion Models with Transformerscited by 6,054
  8. 2022Classifier-Free Diffusion Guidancecited by 5,847
  9. 2022Large Language Models are Zero-Shot Reasonerscited by 4,374
  10. 2023Toolformer: Language Models Can Teach Themselves to Use Toolscited by 4,182
  11. 2021On the Opportunities and Risks of Foundation Modelscited by 4,033
  12. 2024Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameterscited by 1,986
29 more
  1. 2021WebGPT: Browser-assisted question-answering with human feedbackcited by 1,745
  2. 2021An Explanation of In-context Learning as Implicit Bayesian Inferencecited by 938
  3. 2022Transformers learn in-context by gradient descentcited by 732
  4. 2022What learning algorithm is in-context learning? Investigations with linear modelscited by 684
  5. 2016Adaptive Computation Time for Recurrent Neural Networkscited by 506
  6. 2022Analog Bits: Generating Discrete Data using Diffusion Models with Self-Conditioningcited by 373
  7. 2023In-Context Learning Creates Task Vectorscited by 312
  8. 2022Data Distributional Properties Drive Emergent In-Context Learning in Transformerscited by 309
  9. 2023Function Vectors in Large Language Modelscited by 308
  10. 2020Distilling Knowledge from Reader to Retriever for Question Answeringcited by 226
  11. 2023Large Language Models Are Latent Variable Models: Explaining and Finding Good Demonstrations for In-Context Learningcited by 149
  12. 2023We're Afraid Language Models Aren't Modeling Ambiguitycited by 126
  13. 2022Latent Diffusion for Language Generationcited by 115
  14. 2023Clarify When Necessary: Resolving Ambiguity Through Interaction with LMscited by 98
  15. 2023A Theory of Emergent In-Context Learning as Implicit Structure Inductioncited by 86
  16. 2019Meta-learning of Sequential Strategiescited by 70
  17. 2024The broader spectrum of in-context learningcited by 35
  18. 2024Learning Universal Predictorscited by 31
  19. 2025Task Vectors in In-Context Learning: Emergence, Formation, and Benefitcited by 20
  20. 2024Towards an Information Theoretic Framework of Context-Based Offline Meta-Reinforcement Learningcited by 17
  21. 2024Diffusion Guided Language Modelingcited by 15
  22. 2024Does learning the right latent variables necessarily improve in-context learning?cited by 13
  23. 2024Vector-ICL: In-context Learning with Continuous Vector Representationscited by 13
  24. 2024Do LLMs Understand Ambiguity in Text? A Case Study in Open-world Question Answeringcited by 13
  25. 2021Understanding the Origin of Information-Seeking Exploration in Probabilistic Objectives for Controlcited by 10
  26. 2024AmbigNLG: Addressing Task Ambiguity in Instruction for NLGcited by 10
  27. 2023Transformer Language Models Handle Word Frequency in Prediction Headcited by 8
  28. 2024Simulating Weighted Automata over Sequences and Trees with Transformerscited by 4
  29. 2022Beyond Bayes-optimality: meta-learning what you know you don't knowcited by 1

Built on by

Not cited yet.