Lost in the Prompt Order: Revealing the Limitations of Causal Attention in Language Models
Pith reviewed 2026-05-16 12:28 UTC · model grok-4.3
The pith
Causal attention creates an information bottleneck in QOC prompt orders by blocking options from attending to context.
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
The central claim is that the performance superiority of context-first prompts over question-and-options-first prompts stems from the limitations imposed by causal attention. Specifically, when context follows the options in QOC order, the attention mechanism ensures that option tokens receive no information from the subsequent context tokens, leading to poorer decision making across a wide range of models and datasets.
What carries the argument
The causal attention mask, which restricts each token to attend only to preceding tokens in the input sequence.
If this is right
- CQO ordering enables full attention flow from context to options, producing the observed accuracy gains.
- The information bottleneck persists across diverse models and datasets as a direct result of the causal mask.
- Prompt designs that place context after options will systematically underperform on tasks requiring context-informed choices.
- The effect arises from the autoregressive property of the attention mechanism rather than model-specific training.
Where Pith is reading between the lines
- The same bottleneck is likely to appear in chain-of-thought or reading-comprehension tasks where key facts follow the query.
- Hybrid attention patterns that relax the mask for specific prompt sections could reduce order sensitivity without retraining.
- Training on randomly ordered prompts may not fully overcome the limitation because the mask remains fixed at inference time.
Load-bearing premise
The performance gap between CQO and QOC orders is caused primarily by the causal attention mechanism rather than confounding factors such as tokenization differences or training data order.
What would settle it
Modify the causal mask in a controlled experiment to allow option tokens to attend to following context tokens in QOC prompts and measure whether the 14-point accuracy gap with CQO order closes.
read the original abstract
Large language models exhibit surprising sensitivity to the structure of the prompt, but the mechanisms underlying this sensitivity remain poorly understood. In this work, we conduct an in-depth investigation on a striking case: in multiple-choice question answering, placing context before the questions and options (CQO) outperforms the reverse order (QOC) by over 14%p, consistently over a wide range of models and datasets. Through systematic architectural analysis, we identify causal attention as the core mechanism: in QOC prompts, the causal mask prevents option tokens from attending to context, creating an information bottleneck where context becomes invisible to options.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper investigates prompt-order sensitivity in multiple-choice question answering, reporting that context-question-option (CQO) ordering outperforms question-option-context (QOC) ordering by more than 14 percentage points across a range of models and datasets. Through architectural analysis, the authors identify the causal attention mask as the primary cause: in QOC prompts the mask prevents option tokens from attending to preceding context tokens, creating an information bottleneck that renders context invisible to the options.
Significance. If the causal-mask attribution holds after proper isolation from positional and ordering confounders, the result would be significant for understanding fundamental limitations of causal transformers. It would supply a mechanistic account of a large, reproducible prompt-order effect and suggest concrete architectural or inference-time interventions, with direct relevance to prompt engineering and model design.
major comments (2)
- [Abstract] Abstract and the architectural-analysis section: the claim that the CQO-QOC gap is caused by the causal mask creating an information bottleneck is not isolated from positional-embedding and token-order confounders. An ablation that holds the exact token sequence fixed while selectively relaxing only the mask entries between context and option tokens is required; without it the observed gap remains consistent with multiple mechanisms.
- [Results] Results section (performance tables): the reported >14%p gap is presented as consistent, yet no quantitative ablation results or controls for tokenization differences, training-data order statistics, or other architectural elements are described that would rule out alternative explanations for the performance difference.
minor comments (2)
- [Abstract] The abstract states the gap holds 'over a wide range of models and datasets' but does not list the specific models or datasets; adding this information would improve reproducibility.
- [Introduction] Notation for prompt orders (CQO vs. QOC) should be defined explicitly on first use rather than assumed from the title.
Simulated Author's Rebuttal
We thank the referee for the thoughtful and detailed comments. We agree that stronger isolation of the causal-mask effect and additional controls would improve the manuscript. We address each major comment below, indicating the revisions we will make.
read point-by-point responses
-
Referee: [Abstract] Abstract and the architectural-analysis section: the claim that the CQO-QOC gap is caused by the causal mask creating an information bottleneck is not isolated from positional-embedding and token-order confounders. An ablation that holds the exact token sequence fixed while selectively relaxing only the mask entries between context and option tokens is required; without it the observed gap remains consistent with multiple mechanisms.
Authors: We agree that the current experiments do not fully isolate the causal mask from positional and ordering effects. In the revised manuscript we will add a controlled ablation that keeps the exact token sequence and positional embeddings fixed while selectively allowing bidirectional attention only between context and option tokens. Preliminary runs of this ablation recover most of the performance gap, providing direct evidence that the mask is the dominant factor. These results will be reported in a new subsection of the architectural analysis. revision: yes
-
Referee: [Results] Results section (performance tables): the reported >14%p gap is presented as consistent, yet no quantitative ablation results or controls for tokenization differences, training-data order statistics, or other architectural elements are described that would rule out alternative explanations for the performance difference.
Authors: We will expand the results section with quantitative ablations that fix tokenization and prompt formatting across orderings. We have already tested multiple model families and datasets; we will add further breakdowns by model scale and tokenizer type. Controls for training-data order statistics, however, cannot be performed without access to the original pretraining corpora, which are unavailable for the closed models we evaluate. We will explicitly note this limitation and its implications for alternative explanations. revision: partial
- Quantitative controls for training-data order statistics, which would require access to proprietary pretraining data not available to the authors.
Circularity Check
No circularity: empirical gap attributed to standard causal masking properties
full rationale
The paper's derivation rests on direct comparison of CQO vs QOC performance across models and datasets, followed by architectural inspection of the causal mask's effect on attention between context and option tokens. This follows from the fixed definition of causal attention (lower-triangular mask) without any fitted parameters renamed as predictions, without self-citations as load-bearing premises, and without ansatzes or uniqueness theorems imported from prior author work. The information-bottleneck description is a straightforward consequence of the mask definition applied to the QOC token ordering, not a reduction of the observed gap to itself by construction. The analysis is therefore self-contained against external benchmarks of transformer attention mechanics.
Axiom & Free-Parameter Ledger
axioms (1)
- standard math Causal attention mask prevents any token from attending to future tokens in the sequence
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.