Research map / Large language models / Language, safety and interpretability
Where Did It Go Wrong? Attributing Undesirable LLM Behaviors via Representation Gradient Tracing
Zhe Li, Wei Zhao, Yige Li, Jun Sun
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 |
|---|---|---|---|
| replace TracIn with layer-normalised TracIn | TracIn · Harmful data identification (Llama2-7B) | auPRC: 0.117 -> 0.332 | Improved |
| use RepT representation-gradient signatures instead of LESS | LESS · Harmful data identification (Llama2-7B) | auPRC: 0.591 -> 1.000 | Improved |
| use RepT instead of LESS on backdoor detection | LESS · Backdoor poisoning detection (Llama3-8B) | auPRC: 0.048 -> 1.000 | Improved |
| use RepT instead of best baseline LoGra | LoGra · Knowledge contamination Ag->Na (Llama2-7B) | auPRC: 0.588 -> 0.988 | Improved |
| use RepT instead of LESS on full-parameter Llama2-7B | LESS · Harmful data identification (Llama2-7B full-parameter) | P@100: 0.283 -> 0.998 | Improved |
| scale from Llama2-7B to Llama2-70B (LoRA) | TracIn (LN) · Harmful data identification | P@100: 0.683 -> 0.283 | Hurt |
| scale from Llama2-7B to Llama2-70B (LoRA) | RepT · Harmful data identification | P@100: 0.999 -> 0.985 | No clear effect |
| add the representation gradient g_H to the hidden state H in the signature (ablation) | RepT · Harmful data identification | precision | Improved |
| pool all token features instead of last prompt token and first response token (ablation) | RepT · Harmful data identification | precision | Hurt |
| apply Random Projection to 2,048 dimensions (ablation) | RepT · attribution tasks | precision | No clear effect |
| apply Random Shuffle (ablation) | RepT · attribution tasks | precision | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DataInf — DataInf: Efficiently Estimating Data Influence in LoRA-tuned LLMs and Diffusion Models (ICLR 2024) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Language Models Can Predict Their Own Behavior (NeurIPS 2025 · similarity 0.88)
- LLMScan: Causal Scan for LLM Misbehavior Detection (ICML 2025 · similarity 0.88)
- LLMs Know More Than They Show: On the Intrinsic Representation of LLM Hallucinations (ICLR 2025 · similarity 0.88)
- CALM: Curiosity-Driven Auditing for Large Language Models (AAAI 2025 · similarity 0.88)
- Statistical Hypothesis Testing for Auditing Robustness in Language Models (ICML 2025 · similarity 0.88)
- Local Sparsity Enables Unsupervised LLM Safety Detection (NeurIPS 2026 · 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.
