Research map / Large language models / Instruction and fine-tuning
Instructive Decoding: Instruction-Tuned Large Language Models are Self-Refiner from Noisy Instructions
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Instructive Decoding, a training-free contrastive decoding method that contrasts next-token logits from the original instruction against those from noisy/perturbed instructions (e.g., 'opposite'), improving unseen-task generalization of instruction-tuned models.
Problem
Instruction-tuned LMs often fail to follow instructions for unseen tasks outside their training distribution; improving this usually requires costly dataset curation or retraining.
Approach
Adapt contrastive decoding so the 'amateur' signal comes from the same model conditioned on a noisy version of the instruction (truncated/shuffled, null, random words, or a misleading 'opposite' directive), amplifying the anchoring effect toward the true instruction without any parameter updates.
Key results
- With the 'opposite' noisy instruction, Tk-XL overall Rouge-L rises from 45.36 (baseline) to 46.69, and Tk-XXL from 46.01 to 47.43 on the SupNatInst held-out set.
- Tk-XL with ID outperforms its larger sibling Tk-XXL under standard inference on zero-shot Rouge-L, per the paper's claim.
- OpenSNI-7B overall Rouge-L improves from 48.05 to 49.47 with the opposite variant; gains are most notable for Alpaca-7B on UnNatInst.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Opposite noisy instruction in ID | Tk-Large · SupNatInst held-out (zero-shot) | Rouge-L: 41.10 -> 42.21 | Helped |
| Use Opposite noisy instruction in ID | Tk-XL · SupNatInst held-out (zero-shot) | Rouge-L: 45.36 -> 46.69 | Helped |
| Use Opposite noisy instruction in ID | Tk-XXL · SupNatInst held-out (zero-shot) | Rouge-L: 46.01 -> 47.43 | Helped |
| Use Opposite noisy instruction in ID | OpenSNI-7B · SupNatInst held-out (zero-shot) | Rouge-L: 48.05 -> 49.47 | Helped |
| Use ID noisy-instruction variants rather than the standard instruction alone | Tk-Instruct and OpenSNI-7B · SupNatInst held-out (zero-shot), task categories | Rouge-L | Mixed |
| Increase the number of random words in the noisy instruction (ablation) | Tk-Large and Tk-XL · SupNatInst held-out | Rouge-L: 41.77 -> 41.35; 46.46 -> 46.11 | Hurt |
| Vary the Trunc-Shuf truncation ratio (ablation) | Tk-Large and Tk-XL · SupNatInst held-out | Rouge-L: 41.68 to 41.87; 46.21 to 46.67 | Mixed |
| Use sampling-based rather than greedy decoding with null ID (ablation) | Tk-XL · SupNatInst held-out | Rouge-L: 46.35 -> 41.61 ± 0.20 | Hurt |
| Set ID smoothing coefficient epsilon to negative values (ablation) | ID (null) · SupNatInst held-out | Rouge-L | Hurt |
| Set ID smoothing coefficient epsilon above 0.4 (ablation) | ID (null) · SupNatInst held-out | Rouge-L | Hurt |
| Use epsilon between 0.1 and 0.4 (ablation) | ID (null) · SupNatInst held-out | Rouge-L | No clear effect |
| Include two positive demonstration examples with ID (ablation) | ID (null and Opposite) · SupNatInst few-shot | Rouge-L | Mixed |
Surprising findings
- Degradation caused by a noisy instruction in standard decoding correlates positively with the ID improvement it yields; the most harmful 'opposite' instruction gives the largest gain.
- Combining the original instruction with the noisy instruction in the contrast did not improve performance.
Limitations
- Two inference passes per generated output increase computational cost.
- Sensitivity to instruction perturbations and possible error propagation.
- Evaluated mainly on SupNatInst with Rouge-L; sampling decoding degrades performance.
Experimental setup
Datasets: Super-NaturalInstructions UnNatInst
Benchmarks: Super-NaturalInstructions held-out (119 tasks) UnNatInst
Metrics: Rouge-L Exact Match Label Adherence Label Coherence
Baselines: Baseline greedy decoding with original instruction Contrastive Decoding Tk-XXL
Builds on: Contrastive Decoding Super-NaturalInstructions T0 Tk-Instruct
Closest papers by meaning
- Evaluating the Zero-shot Robustness of Instruction-tuned Language Models (ICLR 2024 · similarity 0.91)
- InstructZero: Efficient Instruction Optimization for Black-Box Large Language Models (ICML 2024 · similarity 0.90)
- Enhancing NLU in Large Language Models Using Adversarial Noisy Instruction Tuning (AAAI 2025 · similarity 0.89)
- Instruction Tuning With Loss Over Instructions (NeurIPS 2024 · similarity 0.89)
- LESS: Selecting Influential Data for Targeted Instruction Tuning (ICML 2024 · similarity 0.88)
- InstructRetro: Instruction Tuning post Retrieval-Augmented Pretraining (ICML 2024 · similarity 0.88)
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.