mlsanityData to 7 Oct 2026

In-Context Fixation: When Demonstrated Labels Override Semantics in Few-Shot Classification

arXiv 2605.08295 ↗First posted 8 May 2026cs.LG cs.AI cs.CLcited by 0

Not on the map. Why?

Abstract (arXiv)

While random demonstration labels barely hurt in-context learning (Min et al., 2022), we show that homogeneous labels--even semantically valid ones--collapse accuracy to <=12% across six models (Pythia, Llama, Qwen; 0.8B--8B) and four tasks. The trigger is label-slot content: the model treats tokens occupying the label position as an exhaustive answer vocabulary, with homogeneity as the maximally collapsed case. A novel set-level fixation finding confirms this: when demonstrations carry varied nonsense tokens from {foo,bar,vex,nit,orb}, the model places 42--67% of probability on the demonstrated set while P(dog) remains below 0.2%. This is inconsistent with latent-concept Bayesian accounts (Xie et al., 2022) and reveals that ICL output is constrained vocabulary retrieval--the model binds its output to the demonstrated token inventory regardless of semantic plausibility. The effect generalizes to 4-way classification (0% accuracy across three models, 1B--8B) and multi-token verbalizers ("very positive"), where we decompose fixation into format-level (template adoption) and content-level (polarity override) components that are experimentally dissociable. Mechanistically, per-item paired activation patching on Pythia-1B recovers 98.4% of the gap (95% CI [84%, 112%]), localizing fixation to a layer-7-centered circuit (rank 2/560, 99.8th percentile; 4-fold CV mean 103%). Cross-architecture logit lens on Llama-3.2-1B replicates the encode-then-override trajectory with causal confirmation (top-5 layers: 89% recovery).

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Builds on 42 of its 52 references that are in the map. First posted on arXiv 8 May 2026.

Builds on

42 of 52 references.

References · 42

  1. 2020Language Models are Few-Shot Learnerscited by 32,199
  2. 2024The Llama 3 Herd of Modelscited by 14,192
  3. 2023Lost in the Middle: How Language Models Use Long Contextscited by 3,383
  4. 2022Locating and Editing Factual Associations in GPTcited by 2,609
  5. 2023Pythia: A Suite for Analyzing Large Language Models Across Training and Scalingcited by 1,734
  6. 2022Rethinking the Role of Demonstrations: What Makes In-Context Learning Work?cited by 1,413
  7. 2021Calibrate Before Use: Improving Few-Shot Performance of Language Modelscited by 1,371
  8. 2022Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 smallcited by 1,130
  9. 2021Fantastically Ordered Prompts and Where to Find Them: Overcoming Few-Shot Prompt Order Sensitivitycited by 1,122
  10. 2022In-context Learning and Induction Headscited by 930
  11. 2021An Explanation of In-context Learning as Implicit Bayesian Inferencecited by 915
  12. 2022Transformers learn in-context by gradient descentcited by 714
30 more
  1. 2022What Can Transformers Learn In-Context? A Case Study of Simple Function Classescited by 700
  2. 2023Towards Automated Circuit Discovery for Mechanistic Interpretabilitycited by 700
  3. 2022What learning algorithm is in-context learning? Investigations with linear modelscited by 671
  4. 2023Eliciting Latent Predictions from Transformers with the Tuned Lenscited by 579
  5. 2022Transformer Feed-Forward Layers Build Predictions by Promoting Concepts in the Vocabulary Spacecited by 522
  6. 2023Are Emergent Abilities of Large Language Models a Mirage?cited by 515
  7. 2022Using cognitive psychology to understand GPT-3cited by 389
  8. 2023Larger language models do in-context learning differentlycited by 386
  9. 2023Towards Best Practices of Activation Patching in Language Models: Metrics and Methodscited by 328
  10. 2022Data Distributional Properties Drive Emergent In-Context Learning in Transformerscited by 303
  11. 2023In-Context Learning Creates Task Vectorscited by 302
  12. 2021Surface Form Competition: Why the Highest Probability Answer Isn't Always Rightcited by 256
  13. 2023Label Words are Anchors: An Information Flow Perspective for Understanding In-Context Learningcited by 235
  14. 2023Inverse Scaling: When Bigger Isn't Bettercited by 205
  15. 2022Why Can GPT Learn In-Context? Language Models Implicitly Perform Gradient Descent as Meta-Optimizerscited by 204
  16. 2023Localizing Model Behavior with Path Patchingcited by 182
  17. 2023What In-Context Learning "Learns" In-Context: Disentangling Task Recognition and Task Learningcited by 149
  18. 2023The Hydra Effect: Emergent Self-repair in Language Model Computationscited by 145
  19. 2024In-Context Learning with Long-Context Models: An In-Depth Explorationcited by 126
  20. 2022Ground-Truth Labels Matter: A Deeper Look into Input-Label Demonstrationscited by 116
  21. 2023The mechanistic basis of data dependence and abrupt learning in an in-context classification taskcited by 115
  22. 2023Attention Sorting Combats Recency Bias In Long Context Language Modelscited by 89
  23. 2023Mitigating Label Biases for In-context Learningcited by 88
  24. 2023Overthinking the Truth: Understanding how Language Models Process False Demonstrationscited by 78
  25. 2022Inverse scaling can become U-shapedcited by 76
  26. 2023In-Context Learning Learns Label Relationships but Is Not Conventional Learningcited by 48
  27. 2024Explorations of Self-Repair in Language Modelscited by 35
  28. 2024How do Large Language Models Learn In-Context? Query and Key Matrices of In-Context Heads are Two Towers for Metric Learningcited by 18
  29. 2023How Robust are LLMs to In-Context Majority Label Bias?cited by 13
  30. 2024Does In-Context Learning Really Learn? Rethinking How Large Language Models Respond and Solve Tasks via In-Context Learningcited by 8

Built on by

Not cited yet.