AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

NDAD: Negative-Direction Aware Decoding for Large Language Models via Controllable Hallucination Signal Injection

Panjia Qiu, Mingyuan Fan, Cen Chen, Daixin Wang

ICLR 2026 · Large language models › Language, safety and interpretability · approach: Task vectors and steering

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 changedWhereMeasuredEffect
add NDAD decoding on top of Llama2-7B-base greedyLlama2-7B-base · GSM8K (CoT)accuracy: 13.95 -> 14.86Improved
use NDAD instead of SLEDSLED on Llama2-7B-base · TruthfulQA (MC avg)accuracy: 42.87 -> 43.00No clear effect
use NDAD instead of SLEDSLED on Llama2-7B-base · PopQAexact match: 25.86 -> 26.00No clear effect
use NDAD instead of SLEDSLED on Llama2-70B · GSM8K (CoT)accuracy: 57.01 -> 57.54Improved
use NDAD instead of SLEDSLED on Llama3-8B-instruct · GSM8K (CoT)accuracy: 75.82 -> 77.18Improved
use AD baseline instead of greedy decodingLlama2-7B-base · GSM8K (CoT)accuracy: 13.95 -> 2.12Hurt
select heads randomly instead of by importance (add head-importance selection) (ablation)NDAD · Llama2-7B-base GSM8Kaccuracy: 13.95 -> 14.86Improved
select layers by entropy (add) vs random layer (ablation)NDAD · Llama2-7B-base GSM8Kaccuracy: 14.71 -> 14.86Improved
add global weight (ablation)NDAD · Llama2-7B-base GSM8Kaccuracy: 14.63 -> 14.86Improved
add local weight (ablation)NDAD · Llama2-7B-base GSM8Kaccuracy: 14.10 -> 14.86Improved
add local weight (ablation)NDAD · Llama2-7B-base TruthfulQA (MC avg)accuracy: 42.15 -> 43.00Improved
add local weight (ablation)NDAD · Llama2-7B-base FACTOR (Wiki)accuracy: 61.44 -> 61.57No clear effect

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.

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

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy