mlsanityData to 7 Oct 2026

From Data to Behavior: Predicting Unintended Model Behaviors Before Training

arXiv 2602.04735 ↗First posted 4 Feb 2026cs.LG cs.AI cs.CLcited by 6

Not on the map. Why?

Abstract (arXiv)

Large Language Models (LLMs) can acquire unintended biases from seemingly benign training data even without explicit cues or malicious content. Existing methods struggle to detect such risks before fine-tuning, making post hoc evaluation costly and inefficient. To address this challenge, we introduce Data2Behavior, a new task for predicting unintended model behaviors prior to training. We also propose Manipulating Data Features (MDF), a lightweight approach that summarizes candidate data through their mean representations and injects them into the forward pass of a base model, allowing latent statistical signals in the data to shape model activations and reveal potential biases and safety risks without updating any parameters. MDF achieves reliable prediction while consuming only about 20% of the GPU resources required for fine-tuning. Experiments on Qwen3-14B, Qwen2.5-32B-Instruct, and Gemma-3-12b-it confirm that MDF can anticipate unintended behaviors and provide insight into pre-training vulnerabilities.

Key facts. Cited by 6 papers in the map (data through 7 Oct 2026). Builds on 44 of its 62 references that are in the map. First posted on arXiv 4 Feb 2026.

Builds on

44 of 62 references.

References · 44

  1. 2023Inference-Time Intervention: Eliciting Truthful Answers from a Language Modelcited by 1,259
  2. 2023Representation Engineering: A Top-Down Approach to AI Transparencycited by 1,239
  3. 2023LIMA: Less Is More for Alignmentcited by 1,080
  4. 2023Steering Llama 2 via Contrastive Activation Additioncited by 1,009
  5. 2023Steering Language Models With Activation Engineeringcited by 824
  6. 2024LESS: Selecting Influential Data for Targeted Instruction Tuningcited by 514
  7. 2023Editing Large Language Models: Problems, Methods, and Opportunitiescited by 367
  8. 2025Persona Vectors: Monitoring and Controlling Character Traits in Language Modelscited by 303
  9. 2023From Quantity to Quality: Boosting LLM Performance with Self-Guided Data Selection for Instruction Tuningcited by 301
  10. 2023TRAK: Attributing Model Behavior at Scalecited by 273
  11. 2024A Survey on Data Selection for Language Modelscited by 230
  12. 2024Patchscopes: A Unifying Framework for Inspecting Hidden Representations of Language Modelscited by 225
32 more
  1. 2024Large Language Models for Data Annotation and Synthesis: A Surveycited by 201
  2. 2025AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoderscited by 199
  3. 2025LIMR: Less is More for RL Scalingcited by 136
  4. 2024Superfiltering: Weak-to-Strong Data Filtering for Fast Instruction-Tuningcited by 134
  5. 2024Analyzing the Generalization and Reliability of Steering Vectorscited by 130
  6. 2024What is in Your Safe Data? Identifying Benign Data that Breaks Safetycited by 114
  7. 2024LLM Dataset Inference: Did you train on my dataset?cited by 107
  8. 2024Steering Language Model Refusal with Sparse Autoencoderscited by 89
  9. 2025Subliminal Learning: Language models transmit behavioral traits via hidden signals in datacited by 89
  10. 2024Detoxifying Large Language Models via Knowledge Editingcited by 86
  11. 2024Knowledge Mechanisms in Large Language Models: A Survey and Perspectivecited by 50
  12. 2024LatentQA: Teaching LLMs to Decode Activations Into Natural Languagecited by 49
  13. 2025Massive Values in Self-Attention Modules are the Key to Contextual Knowledge Understandingcited by 44
  14. 2016Auditing Black-box Models for Indirect Influencecited by 41
  15. 2025A Unified Understanding and Evaluation of Steering Methodscited by 40
  16. 2025Inoculation Prompting: Eliciting traits from LLMs during training can suppress them at test-timecited by 32
  17. 2025Language Models are Injective and Hence Invertiblecited by 31
  18. 2025Towards Understanding Subliminal Learning: When and How Hidden Biases Transfercited by 30
  19. 2025We Can't Understand AI Using our Existing Vocabularycited by 25
  20. 2025Fine-Tuning Lowers Safety and Disrupts Evaluation Consistencycited by 25
  21. 2025Narrow Finetuning Leaves Clearly Readable Traces in Activation Differencescited by 24
  22. 2025Beyond Prompt Engineering: Robust Behavior Control in LLMs via Steering Target Atomscited by 23
  23. 2025LogitLens4LLMs: Extending Logit Lens Analysis to Modern Large Language Modelscited by 22
  24. 2025Two Experts Are All You Need for Steering Thinking: Reinforcing Cognitive Effort in MoE Reasoning Models Without Additional Trainingcited by 16
  25. 2025Infini-gram mini: Exact n-gram Search at the Internet Scale with FM-Indexcited by 13
  26. 2025Cross-model Transferability among Large Language Models on the Platonic Representations of Conceptscited by 11
  27. 2025Attack via Overfitting: 10-shot Benign Fine-tuning to Jailbreak LLMscited by 10
  28. 2026Mechanistic Data Attribution: Tracing the Training Origins of Interpretable LLM Unitscited by 10
  29. 2025Speculating LLMs' Chinese Training Data Pollution from Their Tokenscited by 5
  30. 2025Neologism Learning for Controllability and Self-Verbalizationcited by 5
  31. 2025Empirical Evidence for Alignment Faking in a Small LLM and Prompt-Based Mitigation Techniquescited by 4
  32. 2025Subliminal Corruption: Mechanisms, Thresholds, and Interpretabilitycited by 4

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2026: 6.

6 papers.

  1. 2026Subliminal Steering: Stronger Encoding of Hidden Signalscited by 10
  2. 2026Subliminal Learning Is Steering Vector Distillationcited by 9
  3. 2026Shared Latent Structures Enable Unified Backdoor Detection and Mitigation in LLMscited by 0
  4. 2026Towards Identifying the Dataset Biases Causing Phantom Transfercited by 0
  5. 2026Why Ghost Outputs Teach: A Kernel-Based Understanding of Subliminal Learningcited by 0
  6. 2026Alignment Forecasting: Predicting Misalignment From Training Datacited by 0