Research map / Architectures and efficiency / Attention and transformers
Function Induction and Task Generalization: An Interpretability Study with Off-by-One Addition
Qinyuan Ye, Robin Jia, Xiang Ren
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the six function induction heads (versus replacing their outputs with base-run outputs) (ablation) | Gemma-2 (9B) off-by-one addition · 16-shot off-by-one addition, range [0,9] | off-by-one accuracy: 86% -> 0% | Improved |
| keep six randomly chosen heads instead of ablating them (control) (ablation) | Gemma-2 (9B) · 16-shot off-by-one addition | accuracy | No clear effect |
| ablate FI heads with zero or mean ablation instead of instance-specific (ablation) | Gemma-2 (9B) · 16-shot off-by-one addition | contrast and base accuracy: contrast 0%, base 100% | No clear effect |
| add more in-context examples (shots) | six LMs on off-by-one addition · off-by-one addition, range [0,999] | off-by-one accuracy | Improved |
| increase more shots, effect on standard-addition accuracy under contrast prompt | six LMs · off-by-one addition | base (standard addition) accuracy | Hurt |
| enlarge number range from [0,9] to [0,99]/[0,999] | Gemma-2 (9B) and Qwen-2.5 (7B) (digit tokenization) · off-by-one addition | accuracy | Hurt |
| enlarge number range for models with single-token numbers | Llama-2, Mistral, Llama-3, Phi-4 · off-by-one addition | accuracy | No clear effect |
| allow c_test to equal some earlier c_i (no copy restriction) | Mistral-v0.1 (7B), Gemma-2 (9B) · off-by-one addition | accuracy | Improved |
| increase absolute offset k in off-by-k addition | Gemma-2 (9B), Llama-3 (8B) · 32-shot off-by-k addition | accuracy | Hurt |
| ablate FI heads on four task pairs (off-by-two, shifted MMLU, Caesar cipher, base-8) (ablation) | Gemma-2 (9B) · task generalization pairs | contrast accuracy | Hurt |
| identified circuit used alone versus full model | Gemma-2 (9B) · circuit evaluation | faithfulness: (7.17-0.56)/(7.17-(-1.26)) = 78.4% | Improved |
| base-8 addition: model adjustments in no-carry case | Gemma-2 (9B) · Base-8 addition Case 1 | unnecessary adjustment rate: 93% neither; 7% (=6%+1%) adjusted unnecessarily | Mixed |
Closest papers by meaning
- Task Generalization with Autoregressive Compositional Structure: Can Learning from $D$ Tasks Generalize to $D^T$ Tasks? (ICML 2025 · similarity 0.87)
- Language Models Need Inductive Biases to Count Inductively (ICLR 2025 · similarity 0.87)
- Cognitive Mirrors: Exploring the Diverse Functional Roles of Attention Heads in LLM Reasoning (NeurIPS 2025 · similarity 0.87)
- Phenomenal Yet Puzzling: Testing Inductive Reasoning Capabilities of Language Models with Hypothesis Refinement (ICLR 2024 · similarity 0.87)
- Circuit Component Reuse Across Tasks in Transformer Language Models (ICLR 2024 · similarity 0.87)
- Quantifying Generalization Complexity for Large Language Models (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
