mlsanityData to 7 Oct 2026

Learning Concepts, Not Tokens: Self-Supervised Semantic Alignment for Language Models

arXiv 2603.29123 ↗First posted 31 Mar 2026cs.CLcited by 1

On the map: LLM alignment / RLHF / preference optimisation → LLM preference optimization methods

Abstract (arXiv)

The next-token prediction (NTP) objective trains language models to predict a single token at each step, even though many continuations can express the same meaning. For example, in the sentence ``this sticker can be placed here'', positioned, attached, or put are all plausible alternatives. While standard NTP training treats these alternatives as mutually exclusive targets, we explore a self-supervised framework that encourages models to predict concepts, approximated as sets of semantically equivalent tokens. Models trained with this concept supervision align better with human similarity judgments, improve classification, clustering, and reranking performance, and achieve comparable or stronger downstream reasoning. These gains come with lower perplexity on semantically meaningful words (Section 3.2) and only minimal increases in global perplexity, suggesting that concepts enhance semantic alignment while preserving language modeling quality. Our code is available at https://github.com/christine-zhang1/learning-concepts

Key facts. Cited by 1 paper in the map (data through 7 Oct 2026). Placed in LLM preference optimization methods, one of 10 directions in LLM alignment / RLHF / preference optimisation. Builds on 51 of its 78 references that are in the map. First posted on arXiv 31 Mar 2026.

Builds on

51 of 78 references.

Closest ancestors · 2

Same direction, newest first

  1. 2023LIMA: Less Is More for Alignment18 May
  2. 2022Training language models to follow instructions with human feedback4 Mar

Other references · 49

  1. 2019Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformercited by 14,491
  2. 2019Sentence-BERT: Sentence Embeddings using Siamese BERT-Networkscited by 9,049
  3. 2023Direct Preference Optimization: Your Language Model is Secretly a Reward Modelcited by 7,916
  4. 2018Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challengecited by 4,710
  5. 2023Mamba: Linear-Time Sequence Modeling with Selective State Spacescited by 4,616
  6. 2019HellaSwag: Can a Machine Really Finish Your Sentence?cited by 4,018
  7. 2023QLoRA: Efficient Finetuning of Quantized LLMscited by 3,355
  8. 2019PIQA: Reasoning about Physical Commonsense in Natural Languagecited by 2,818
  9. 2017A Broad-Coverage Challenge Corpus for Sentence Understanding through Inferencecited by 2,665
  10. 2021SimCSE: Simple Contrastive Learning of Sentence Embeddingscited by 2,168
  11. 2018Can a Suit of Armor Conduct Electricity? A New Dataset for Open Book Question Answeringcited by 2,114
  12. 2015A large annotated corpus for learning natural language inferencecited by 1,837
37 more
  1. 2025Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Modelscited by 1,175
  2. 2022Diffusion-LM Improves Controllable Text Generationcited by 1,006
  3. 2017SemEval-2017 Task 1: Semantic Textual Similarity - Multilingual and Cross-lingual Focused Evaluationcited by 998
  4. 2024NV-Embed: Improved Techniques for Training LLMs as Generalist Embedding Modelscited by 466
  5. 2020Efficient Intent Detection with Dual Sentence Encoderscited by 419
  6. 2023Fine-Grained Human Feedback Gives Better Rewards for Language Model Trainingcited by 399
  7. 2024LLM2Vec: Large Language Models Are Secretly Powerful Text Encoderscited by 396
  8. 2021Surface Form Competition: Why the Highest Probability Answer Isn't Always Rightcited by 256
  9. 2022SGPT: GPT Sentence Embeddings for Semantic Searchcited by 204
  10. 2020MTOP: A Comprehensive Multilingual Task-Oriented Semantic Parsing Benchmarkcited by 156
  11. 2024The pitfalls of next-token predictioncited by 156
  12. 2025MMTEB: Massive Multilingual Text Embedding Benchmarkcited by 120
  13. 2021TSDAE: Using Transformer-based Sequential Denoising Auto-Encoder for Unsupervised Sentence Embedding Learningcited by 116
  14. 2025Soft Thinking: Unlocking the Reasoning Potential of LLMs in Continuous Concept Spacecited by 109
  15. 2024Large Concept Models: Language Modeling in a Sentence Representation Spacecited by 86
  16. 2025PlanGenLLMs: A Modern Survey of LLM Planning Capabilitiescited by 82
  17. 2016Reward Augmented Maximum Likelihood for Neural Structured Predictioncited by 80
  18. 2019SenseBERT: Driving Some Sense into BERTcited by 73
  19. 2017A Continuously Growing Dataset of Sentential Paraphrasescited by 72
  20. 2021I Wish I Would Have Loved This One, But I Didn't -- A Multilingual Dataset for Counterfactual Detection in Product Reviewscited by 44
  21. 2021TWEAC: Transformer with Extendable QA Agent Classifierscited by 37
  22. 2025LLM-JEPA: Large Language Models Meet Joint Embedding Predictive Architecturescited by 35
  23. 2018Adversarial Domain Adaptation for Duplicate Question Detectioncited by 30
  24. 2025From Tokens to Thoughts: How LLMs and Humans Trade Compression for Meaningcited by 23
  25. 2024TEncDM: Understanding the Properties of the Diffusion Model in the Space of Language Model Encodingscited by 22
  26. 2025SofT-GRPO: Surpassing Discrete-Token LLM Reinforcement Learning via Gumbel-Reparameterized Soft-Thinking Policy Optimizationcited by 18
  27. 2026ProFit: Leveraging High-Value Signals in SFT via Probability-Guided Token Selectioncited by 16
  28. 2025Beyond Multi-Token Prediction: Pretraining LLMs with Future Summariescited by 14
  29. 2020Semantic Label Smoothing for Sequence to Sequence Problemscited by 13
  30. 2025I Predict Therefore I Am: Is Next Token Prediction Enough to Learn Human-Interpretable Concepts from Data?cited by 13
  31. 2022Always Keep your Target in Mind: Studying Semantics and Improving Performance of Neural Lexical Substitutioncited by 12
  32. 2025Efficient Joint Prediction of Multiple Future Tokenscited by 10
  33. 2024Can You Learn Semantics Through Next-Word Prediction? The Case of Entailmentcited by 9
  34. 2018Token-level and sequence-level loss smoothing for RNN language modelscited by 6
  35. 2026Beyond Tokens: Concept-Level Training Objectives for LLMscited by 3
  36. 2025Not quite Sherlock Holmes: Language model predictions do not reliably differentiate impossible from improbable eventscited by 2
  37. 2025Improving Large Language Models with Concept-Aware Fine-Tuningcited by 2

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2026: 1.

1 paper.

  1. 2026From Found to Designed: Concepts as a Design Axis for Large Language Modelscited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2023Text Alignment Is An Efficient Unified Model for Massive NLP Tasks
  2. 2026Soft Token Alignment for Cross-Lingual Reasoning
  3. 2025TokAlign: Efficient Vocabulary Adaptation via Token Alignment
  4. 2024Self-Supervised Alignment with Mutual Information: Learning to Follow Principles without Preference Labels
  5. 2026TokAlign++: Advancing Vocabulary Adaptation via Better Token Alignment
  6. 2025Alternatives To Next Token Prediction In Text Generation -- A Survey
  7. 2021Improving Pretrained Cross-Lingual Language Models via Self-Labeled Word Alignment
  8. 2025Align-then-Unlearn: Embedding Alignment for LLM Unlearning