Research map / Large language models / Instruction and fine-tuning

Order-Independence Without Fine Tuning

NeurIPS 2024 · Large language models › Instruction and fine-tuning · approach: Fine-tuning generalization and delta compression · 1 citations (OpenAlex)

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 Set-Based Prompting, an inference-time change to attention mask and positional encoding that makes parallel sub-sequences (e.g., MCQ options) order-independent for pretrained transformer LLMs without fine-tuning.

Problem

LLM outputs change significantly when semantically identical sub-sequences (such as multiple-choice options or documents) are reordered, causing order dependency that harms reliability and comparability of benchmark results.

Approach

Remove ordering information between a specified set of sub-sequences by using a non-triangular attention mask (tokens see only their own sub-sequence, shared prefix, and label-free tokens) together with positional indices reset at the start of each sub-sequence, provably making outputs invariant to sub-sequence permutations, applied drop-in at inference time with no training.

Key results

• Set-Based Prompting yields identical outputs under all option orderings; for most models its accuracy falls within the range spanned by normal vs reversed ordering on CSQA and MMLU (exception: Llama 3 8B Instruct).
• On MMLU, for all Llama 2 models Set-Based Prompting lies above or within the interquartile range of 24 permutations; Llama 3 is as bad as the worst ordering but still within range.
• Llama 2 7B summarized a 6,018-token Wikipedia article when its sections were fed in parallel, whereas sequential input produced degenerate repeated text.

Measured effects reported in the paper

Surprising findings

• Modifying only the attention mask reduces order variation more than modifying only the positional encoding, though the effect varies by model.
• Adding enumerated option numbers improved accuracy across orderings, but this reintroduces order information and removes the invariance guarantee.

Limitations

• Out-of-distribution activations lower accuracy relative to the unmodified model, especially for some models (e.g., Llama 3 8B Instruct), and no fine-tuning was used to adapt models to the method.
• Evaluated only single parallel sub-sequence per prompt and four model families; each family needs custom inference code.
• Results on MMLU/CSQA use a modified prompt format (quoted options) and a 1-shot setup, so absolute accuracies differ from standard benchmarks; 70B models not tested.

Experimental setup

Datasets: CommonsenseQA MMLU Financial Phrase Bank Long-context retrieval (Liu et al. setup)

Benchmarks: CommonsenseQA MMLU Financial Phrase Bank sentiment ICL long-context document retrieval

Metrics: accuracy worst-of-2 accuracy best-of-2 accuracy accuracy distribution across all 24 MMLU permutations

Baselines: unmodified model with normal ordering unmodified model with reversed ordering partial implementations (attention mask only, positional encoding only)

Builds on: transformer self-attention rotary positional embeddings absolute positional embeddings

Closest papers by meaning

• Set-LLM: A Permutation-Invariant LLM (NeurIPS 2025 · similarity 0.90)
• Large Language Diffusion Models (NeurIPS 2025 · similarity 0.87)
• Guaranteed Generation from Large Language Models (ICLR 2025 · similarity 0.87)
• Beyond Autoregression: Fast LLMs via Self-Distillation Through Time (ICLR 2025 · similarity 0.87)
• The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models (ICML 2026 · similarity 0.86)
• Large language models suffer from a curse of ambiguity (NeurIPS 2026 · similarity 0.86)