Counterfactual Training Improves LLM Self-Attribution of Sources
Zikai Zhou · hf · 2026-10-28
A paper released on Hugging Face introduces CREDIT, a counterfactual training method that improves language models' self-attribution of external information, reducing false acknowledgment of sources that did not actually influence the output.
- Problem: models often cannot accurately report which external sources they used.
- Method: counterfactual training with paired samples where removing a source should change the output.
- Result: fewer false acknowledgments of non-influential sources.
More from Research
- DeepMind researcher: all of mathematics will soon be formalizable, with agents hunting proofs — ChrSzegedy · 2026-10-28
- A fine-tuned 'cat obsession' control fools every AI welfare criterion — MirrorEthic_Anchor · 2026-10-28
- Sweeping every Qwen3-8B neuron, rated for weirdness by Haiku — Sauers_ · 2026-10-28
- Kevin Buzzard: Lean formalization revealed errors in theorems by 4 Fields medalists — Machine Learning Street Talk · 2026-10-28
- Follow-up repo machine-verifies Hamiltonian cycle results on top of OpenAI/Anthropic ETH refutation claims — xennygrimmato_ · 2026-10-28
- DiscoTrace wins EMNLP 2026 Best Paper: LLM answers lack humans' community-specific rhetorical diversity — TuhinChakr · 2026-10-28