Source-linked AI summary
In-Place Instruction Following in Diffusion Language Models
Zheng Nie, Zherui Li, Jiaming Zhang, Kun Wang, Zhenhong Zhou, Yufei Guo
TL;DR
Prefix-based, left-to-right prompting is not native to arbitrary-position constraints, which arise in scenarios requiring local control. The paper formalizes In-place Instruction Following, builds IIF-Bench, and proposes Graft, a post-training framework. Across four representative dLLMs, Graft raises the average overall IIF score by 15.35 points, from 57.75 to 73.10.
Problem
Prefix-based, left-to-right prompting is not native to arbitrary-position constraints, though some scenarios require constraints at specific output positions.
Source: §1 · p. 1Method
Results
15.35 points: Graft raises the average overall IIF score from 57.75 to 73.10 on four representative dLLMs.
Source: §1 · p. 2Takeaways & Limitations
An attention-bias probe finds evidence that insufficient prioritization of constraint spans is one source of vanilla dLLM failures during parallel denoising.
Source: §7 · p. 8Evaluation covers four representative dLLMs; scalability to larger diffusion models and robustness under more complex or conflicting anchors require further investigation.
Source: Limitations · p. 9
Abstract
from arXiv · showhide
Diffusion Large Language Models (dLLMs) generate text via bidirectional iterative denoising, naturally supporting user-specified constraints anchored at arbitrary output positions, a paradigm known as In-place Prompting (IPP). We formalize this as the In-place Instruction Following (IIF) task and construct IIF-Bench, a hierarchical benchmark spanning literal, style, and discourse-function constraints, paired with a rubric-based local-global evaluation protocol. An inference-time attention-bias probe suggests that vanilla dLLMs often under-prioritize constraint spans during denoising. We then propose GRAFT, an IPP-oriented post-training framework combining constraint-aware SFT and preference optimization. On four representative dLLMs, GRAFT raises the average IIF score from 57.75 to 73.10 (+15.35 points), with absolute gains of 15.91 and 15.57 points on literal and discourse-function constraints, while preserving general generation ability.
1 Introduction
The paper frames in-place constraints as a control capability for diffusion LLMs, formalizes and benchmarks that task, and proposes Graft to improve constraint following while balancing local and global quality.
1 Introduction
- Diffusion LLMs can anchor constraints at arbitrary output positions, but reliable execution alongside natural transitions and global semantic consistency remains an open question.The paper contrasts this control interface with prefix-based prompting and the additional constrained decoding, infilling training, or multi-round editing often needed for arbitrary-position control in autoregressive models.Source: §1 · p. 1
- IIF-Bench evaluates literal, style, and discourse-function constraints with a rubric-based local-global protocol that assesses execution, boundary coherence, global consistency, and fluency.Automatic judgments are cross-validated with a specialized reward model and human evaluation.Source: §1 · p. 2
- The attention-bias probe finds that moderate bias improves constraint following, while excessive bias harms generation quality, motivating Graft’s post-training approach.Graft combines constraint-aware SFT, which treats in-place spans as fixed conditioning variables, with preference optimization for local execution, boundary coherence, and global consistency.Source: §1 · p. 2
- Graft raises the average overall IIF score from 57.75 to 73.10 (+15.35 points) across four representative dLLMs.Absolute gains are 15.91 points on literal constraints and 15.57 points on discourse-function constraints; local execution, boundary coherence, and global consistency improve with only minor impact on general generation ability.Source: §1 · p. 2
- Graft is presented as an IPP-oriented post-training framework for improving in-place constraint following while mitigating conflict between local constraints and global consistency.The paper also identifies fine-grained evaluation and an attention-bias probe as tools for analyzing a key bottleneck in dLLMs.Source: §1 · p. 2
2 Background
Standard dLLM decoding appends masked blocks to the end of context and resolves them through iterative denoising. IPP instead places a fixed anchor between two generated spans, which the model fills bidirectionally.
2.1 Inference Paradigm of dLLMs
- Standard dLLMs decode text in blocks, selecting high-confidence masked positions for prediction and re-masking unselected positions until the block is resolved.Resolved blocks are merged into context before generation proceeds to the next block.Source: §2.1 · p. 2
- In conventional decoding, each masked block is appended to the end of context, making the dLLM a semi-autoregressive generator.Figure 1 compares this standard paradigm with IPP.Source: §2.1 · p. 3, Figure 1 · p. 3
2.2 Inference Paradigm of IPP
- IPP divides the target into a prompt, front span, immutable anchor, and back span, then denoises only the generated spans while keeping the prompt and anchor fixed.The fixed anchor conditions bidirectional generation around it.Source: §2.2 · p. 3
3 IIF-Bench
IIF-Bench evaluates whether diffusion LLMs can follow constraints placed within an output while maintaining local coherence and global consistency. It spans three constraint levels and uses a fine-grained local-global rubric.
3.1 Motivation and Task Formulation
- IIF-Bench targets in-output constraints and the tension between localized control and global coherence, extending evaluation beyond prefix constraints and easily verified surface forms.Its three-level difficulty spectrum runs from literal to discourse constraints.Source: §3.1 · p. 3
- IIF requires generating the spans around a fixed, tagged control span so the constraint is realized while the output remains locally coherent and globally consistent with the user instruction.The task is expressed as generating ygen conditioned on the prompt and anchor subject to a level-specific realization criterion.Source: §3.1 · pp. 3–4
3.2 Dataset Introduction
- IIF-Bench contains approximately 5K instances from sources including Alpaca, LFQA, and Poetry, organized into literal, style, and discourse-constraint levels.The levels test exact phrase embedding, realization of one of 12 literary styles, and execution of eight logical actions, respectively.Source: §3.2 · p. 4
3.3 Multi-dimensional Evaluation Protocol
- The evaluation uses four rubric dimensions—local execution, contextual integration, boundary quality, and global requirement—with level-specific criteria for literal and style constraints.For discourse constraints, the rubric also checks whether the required function is performed, aligns with the global argument, has a clear boundary, and is grammatically clear.Source: §3.3 · pp. 4–5
- Rubric-based LLM-as-judge evaluation uses GPT-5 as the primary evaluator and JudgeLM as an additional evaluator, with human evaluation details in the appendix.Source: §3.3 · p. 5
4 Method
The method probes whether dLLMs under-prioritize in-place constraints during denoising, then proposes two-stage Graft post-training to improve constraint following while maintaining global consistency.
4.1 Attention Bias Probe
- Moderate attention bias improves constraint adherence and overall quality, while excessive bias harms fluency and contextual fit.The probe suggests vanilla dLLMs often assign the constraint span insufficient priority rather than lacking the linguistic capability to follow it.Source: §4.1 · p. 5
4.2 In-place Oriented Post-Training
- Graft aims to internalize the probe’s inference-time attention bias through IPP-specific post-training, treating the user-inserted anchor as a fixed constraint rather than a generation target.The framework has two stages: Graft-SFT and Graft-DPO.Source: §4.2 · pp. 5–6
4.2.1 Graft-SFT
- Graft-SFT treats the anchor as fixed conditioning and masks loss to generated front and back spans rather than predicting the constraint.This localized masking is intended to focus parameter updates on harmonizing surrounding context with the constraint.Source: §4.2.1 · p. 6
4.2.2 Graft-DPO
- Graft-DPO uses preference pairs favoring responses that balance anchor adherence with global consistency, and evaluates the preference objective only on generated spans.The anchor remains a grounded condition during preference optimization.Source: §4.2.2 · p. 6
5 Experiments
Experiments compare four diffusion LLMs and autoregressive baselines on IIF-Bench, then assess overall performance, denoising changes, and test-time ablations. Graft-aligned models improve IIF scores, selectively alter anchor-related representations, and reduce reliance on some test-time interventions.
5.1 Setup
- Experiments use four dLLM backbones on IIF-Bench, compare vanilla models and prompt-adapted autoregressive baselines, and average GPT-5 and JudgeLM scores over three runs.The autoregressive baselines receive natural-language prompts describing the anchor, constraint type, and intended position.Source: §5.1 · p. 6
5.2 Overall Performance
- Graft-aligned dLLMs improve overall IIF scores by over 15 points on average and outperform vanilla counterparts at every constraint level.GPT-4o slightly leads at L3, while Graft-aligned models lead at L1 and L2.Source: §5.2 · p. 6
- Graft-aligned models maintain robust general capabilities, with negligible impact on general generation ability.Figure 5 compares LLaDA and Dream across general benchmarks.Source: §5.2 · p. 7, Figure 5 · p. 7
5.4 Ablation Study
- Moderate frequency penalties perform best, and Graft-aligned models are less sensitive at low penalties; after alignment, tag-suppression gains narrow as formatting boundaries are internalized.Excessive frequency penalties degrade both model variants by suppressing necessary literal phrases and discourse connectors.Source: §5.4 · p. 7
5.3 Attention Diagnosis
- Graft selectively strengthens anchor-related representations: LLaDA shows its largest positive change in upper layers, while Dream’s positive change extends into middle layers.The layer-step maps show that Graft does not overwrite the whole pretrained denoising process.Source: Figure 6 · p. 7, §5.3 · p. 7
6 Related Work
Related work situates this study in masked diffusion language modeling and emerging mid-sequence control. The paper distinguishes its focus on user-facing localized constraints that must remain globally coherent.
Diffusion Language Models
- Masked diffusion models use bidirectional iterative refinement, providing a modeling basis for infilling and arbitrary-position control.Prior work spans continuous denoising, discrete token corruption, and scalable masked diffusion models.Source: §6 · pp. 7–8
In-place Prompting and Mid-sequence Control
- Prior mid-sequence methods insert reasoning, templates, or examples during denoising; this work instead studies user-facing localized constraints under global coherence.Related safety work also examines diffusion-specific attack surfaces.Source: §6 · p. 8
7 Conclusion
The paper formalizes in-place instruction following and builds a benchmark spanning three difficulty levels. It uses an attention-bias probe to diagnose constraint prioritization and proposes Graft post-training, while noting a persistent local–global trade-off for style constraints.
7 Conclusion
- The paper formalizes In-Place Instruction Following and constructs a benchmark spanning three difficulty levels.The benchmark frames instruction following around constraints placed within generated outputs.Source: §7 · p. 8
- An attention-bias probe finds evidence that vanilla dLLMs sometimes under-prioritize constraint spans during parallel denoising.Source: §7 · p. 8
- Graft combines Graft-SFT and Graft-DPO, while fine-grained evaluation finds a persistent local–global consistency trade-off in style constraints.Source: §7 · p. 8
Limitations
The paper identifies limits in benchmark coverage, rubric-based evaluation, and the scope of Graft’s evaluation. It also states that language models were used only to polish the manuscript.
Limitations
- IIF-Bench focuses mainly on text-only following, with limited coverage of multilingual, multimodal, and interactive editing scenarios.Source: Limitations · p. 9
- Rubric-based evaluation may be affected by evaluator bias, especially for style and discourse-function constraints.Source: Limitations · p. 9
- Graft was evaluated on four representative dLLMs; scalability to larger models and robustness to complex or conflicting anchors remain open questions.Source: Limitations · p. 9
A LLM Usage Statement
- Language models were used to refine and polish the manuscript, not to generate new content or ideas.Source: A LLM Usage Statement · p. 11
B Experimental Details
The experiments evaluate Graft on partially constrained continuation tasks, comparing diffusion models under native in-place prompting with autoregressive baselines prompted to satisfy the same positional constraints.
Models and inference
- Four diffusion backbones—LLaDA-1.5, LLaDA-8B, Dream-7B, and MMaDA—are evaluated with the inserted constraint fixed in the output canvas.Vanilla diffusion models use the same IPP format as Graft-aligned variants; the generation length is instance-specific and capped by a 1024-token full-sequence limit.Source: B Experimental Details · pp. 11–12, Table 5 · p. 12
Benchmark construction
- IIF-Bench converts instruction–response pairs into partially constrained continuations spanning literal, semantic-style, and discourse-function constraints.Literal examples use exact sentences or verbatim keywords; style examples rewrite continuations in one of twelve styles, and discourse examples use one of eight functions.Source: B Experimental Details · p. 11
Training and preference data
- Training uses disjoint IIF-instance splits; Graft-SFT targets constraint-satisfying rewrites, while Graft-DPO pairs them against unconstrained or pre-rewrite responses.The data construction is applied across L1 literal, L2 style, and L3 discourse-function constraints.Source: B Experimental Details · p. 12
Baselines and benchmark statistics
- Autoregressive baselines generate text from left to right and are prompted to produce a complete response satisfying the same in-place constraint specification used for the comparison.The baselines include GPT-4o, Llama-3.1-8B-Instruct, and Qwen3-8B-Instruct. Table 4 reports constraint-length statistics across the three levels of the IIF benchmark.Source: B Experimental Details · p. 12, Table 4 · p. 12
C Extended Experimental Analysis
Extended analyses find that anchor placement and quantity affect Graft’s performance: central single anchors fare best, while multiple anchors expose a coordination boundary.
Ablation on Anchor Positions
- Late-position L2 performance diverges: LLaDA improves, while Dream degrades.The authors link the difference to LLaDA’s block-wise refinement of preceding context and Dream’s attenuated control signals over an expanded attention radius.Source: C Extended Experimental Analysis · p. 13
- Overall scores peak with middle-positioned single anchors for both LLaDA-8B and Dream-7B.Early, middle, and late positions correspond to 10%–30%, 40%–60%, and 70%–90% of the sequence; the authors attribute the central advantage to balanced bidirectional context.Source: Table 8 · p. 13, C Extended Experimental Analysis · p. 13
Scaling of Multiple Anchors
- With multiple anchors, competing constraints fragment attention; literal constraints remain the most resilient, while local execution and global coherence become difficult to maintain.The authors attribute the difficulty to attention priorities calibrated for a single focal point.Source: C Extended Experimental Analysis · p. 14
- Moving from one to two uniformly distributed anchors causes a severe performance fracture, especially for L3 discourse-function constraints.The decline saturates sub-linearly as anchors increase; the authors identify multi-focal coordination as a generalization boundary for single-anchor post-training.Source: C Extended Experimental Analysis · pp. 13–14, Table 9 · p. 14
D Human Evaluation Details
The human-evaluation details pair per-dimension assessment with rubrics for local execution and broader coherence, including a conflict adjustment for stylistic tensions and qualitative case studies.
D Human Evaluation Details
- The double-blind human study samples IIF-Bench examples and reports per-dimension breakdowns across all three difficulty levels.Table 11 complements the main text’s holistic scores with fine-grained human-evaluation results.Source: D Human Evaluation Details · p. 14, Table 11 · p. 15
E Evaluation Rubrics
- The rubric separates local constraint execution, contextual integration, boundary quality, and global requirement into D1–D4.These structural labels apply across levels, but their sub-criteria differ for literal fidelity, semantic realization, and discourse-function execution.Source: E Evaluation Rubrics · p. 14
- Rubric totals sum D1–D4 to a maximum of 100, while L3-D1 criteria vary by discourse function.Tables 12 and 13 provide per-sub-item point allocations and function-specific L3-D1 instantiations.Source: E Evaluation Rubrics · p. 14, Table 12 · p. 18, Table 13 · p. 19
- For L2-D2, the conflict adjustment accounts for register distance, emotional-valence conflict, and structural incompatibility when judging contextual integration.The adjustment is compensatory: higher conflict reduces the penalty from a low raw D2 score, with β controlling adjustment strength.Source: E Evaluation Rubrics · pp. 14–15
- The reported L2-D2 value is the mean adjusted score across L2 instances, and the overall L2 score sums D1raw, Sadj, D3raw, and D4raw.Because Sadj is averaged across instances, the reported L2-D2 value is not restricted to the discrete annotation grid.Source: E Evaluation Rubrics · p. 15
F Case Study
- In the style case, Graft maintains the requested bright pastoral tone more consistently, whereas Vanilla opens with imagery that conflicts with it.Graft uses green fields, warm sunlight, birdsong, and streams; Vanilla later partially recovers after an opening of twilight, shadows, and moonlight.Source: F Case Study · p. 15, Figure 7 · p. 16
- In the discourse case, Graft contests a plausible imaging assumption with relevant pipeline components, while Vanilla’s gravity-and-pressure reference is distracting and technically weak.Graft identifies electron optics and detector systems as necessary parts of the imaging pipeline.Source: F Case Study · p. 15, Figure 8 · p. 17