Precise instruction-following is a fundamental ability of large language models (LLMs), requiring their outputs to strictly satisfy objective constraints in input instructions. In complex application scenarios, these constraints often possess diverse scopes that govern specific response segments rather than the entire output. However, existing optimization methods often neglect constraint scope during data construction and rely on binary per-constraint rewards, yielding limited data diversity and sparse supervision for complex constraints. To this end, we propose ScopeIF, a novel training framework for scope-aware precise instruction-following. We first introduce a unified schema that factorizes objective constraints into three decoupled dimensions: Scope, Target, and Range. Grounded in this schema, we construct ScopeInstruct, a large-scale instruction dataset with diverse scope-aware constraints, and combine tool-grounded verification with graded reward modeling to quantify the violation degree of each constraint, providing dense supervision for policy optimization. Extensive experiments demonstrate that ScopeIF consistently outperforms existing methods, particularly on complex scope-aware constraints, while preserving general capabilities. Notably, it enables optimized Qwen3-4B and 8B models to rival or surpass strong frontier models such as Gemini-2.5-Pro and DeepSeek-V3.2, establishing an effective paradigm for advancing scope-aware instruction-following. Our code and data are available at https://github.com/thu-coai/ScopeIF.
Figures & tables
Figure 1: An example of instruction with scope-aware constraints that govern specific response segments rather than the entire response.
Figure 2: Overall framework of ScopeIF. Top: Complex instruction collection via factorized constraint synthesis. Bottom: Policy model optimization via graded reward modeling.
Model
ScopeIF-Test
IFEval
IFBench
IFHierBench
Avg.
ISR
CSR
P-S
P-L
I-S
I-L
P-S
P-L
I-S
I-L
P-S
I-S
Frontier Models
Seed-2.0-Pro
21.8
69.5
88.2
91.7
92.3
94.6
49.3
56.0
51.5
57.6
22.8
52.8
57.2
Gemini-2.5-Pro
15.8
63.7
90.0
93.0
92.8
95.2
40.0
47.7
43.0
50.9
20.8
46.7
52.9
DeepSeek-V3.2
9.1
50.8
87.4
91.5
91.2
94.0
38.3
43.3
42.2
47.1
14.3
33.7
46.9
GPT-4.1
8.6
53.1
87.8
91.3
91.5
94.4
36.7
41.3
41.3
46.2
18.2
44.1
48.7
Table 1: Experimental results (%) of different optimization methods on instruction-following benchmarks. P and I stand for prompt-level and instruction-level metrics, respectively. L and S denote loose and strict evaluations, respectively. HRA means hierarchical reward aggregation. The highest result for each backbone model is bolded , and the overall highest is underlined .
Model
Method
#Data
ScopeIF-Test
IFEval
IFBench
IFHierBench
Avg.
ISR
CSR
P-S
P-L
I-S
I-L
P-S
P-L
I-S
I-L
P-S
I-S
HIR
16k
6.7
44.2
87.2
91.0
90.4
92.9
41.7
44.7
44.2
47.7
16.0
25.4
45.3
VerIF
22k
7.9
47.7
86.5
89.8
91.0
93.3
42.3
48.3
45.3
50.9
16.5
30.9
47.1
MulDimIF
8k
6.0
42.6
85.6
88.5
90.4
92.3
39.3
43.0
41.6
46.5
14.7
21.6
43.6
Qwen3-4B
ScopeIF (Ours)
16k
15.3
60.9
88.2
90.9
92.2
94.1
49.7
53.3
52.0
56.1
18.8
39.4
52.8
HIR
16k
8.6
48.9
89.5
92.1
92.7
94.6
39.3
44.3
41.9
46.8
17.3
30.7
47.0
Table 2: Experimental results (%) of different instruction-following improving frameworks. To ensure fairness in the amount of training data, MulDimIF used twice the number of training epochs.
Figure 3: Performance of constraints with different scope categories on our test set.
Dataset
Self-BLEU-4 ↓
Distinct-2 ↑
Entropy-4 ↑
HIR
0.931
0.067
10.3
VerIF
0.918
0.068
11.0
MulDimIF
0.961
0.059
7.8
ScopeInstruct
0.834
0.133
12.8
Table 5: Constraint diversity of different training datasets for instruction-following.
Figure 4: Average performance of IFEval, IFBench, and IFHierBench (a), and the proportion of rollout response pairs with identical rewards (b) during ScopeIF training.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: The Scope (a) and Range (b) distribution of constraints in ScopeInstruct .
Figure 6: The Target distribution of constraints in ScopeInstruct .
Figure 7: Mean squared distance from each constraint embedding to its assigned k -means centroid under varying numbers of clusters k .
Figure 8: Distribution of the number of constraints per instruction in ScopeInstruct .
Model
Global
Traversal
Position
Condition
Nested
Frontier Models
Seed-2.0-Pro
75.1
59.8
70.8
63.5
64.9
Gemini-2.5-Pro
71.5
56.6
60.8
61.0
56.9
DeepSeek-V3.2
61.5
45.7
45.1
44.2
43.1
GPT-4.1
62.5
49.2
49.0
46.3
45.6
QwQ-32B
58.3
40.4
45.5
43.3
39.9
Appendix
Table 6: CSR (%) across five Scope categories on our test set. The highest result for each backbone model is bolded , and the overall highest is underlined .
Model
Capacity
Literals
Structure
Checklist
Pattern
Relations
Frontier Models
Seed-2.0-Pro
61.4
69.5
79.1
74.2
74.6
58.2
Gemini-2.5-Pro
55.5
64.6
73.8
68.3
69.8
51.5
DeepSeek-V3.2
43.8
52.2
63.4
52.6
57.2
35.2
GPT-4.1
45.6
51.2
69.6
55.7
61.8
37.3
QwQ-32B
39.1
47.6
64.3
52.7
53.7
36.6
Appendix
Table 7: CSR (%) across six Target categories on our test set.
Model
Absolute Range
Relative Range
One-Sided Bounds
Prohibition
Equality
Intervals
Comparison
Arithmetic Relations
Frontier Models
Seed-2.0-Pro
80.3
82.0
61.0
61.5
68.9
40.2
Gemini-2.5-Pro
76.8
80.9
50.2
56.1
70.6
32.2
DeepSeek-V3.2
68.1
74.2
34.5
41.8
47.9
15.0
GPT-4.1
70.8
76.3
37.0
41.2
57.0
16.3
Appendix
Table 8: CSR (%) across six Range categories on our test set.
Instruction-following ability is critical for deploying large language models in real-world applications, where downstream components depend on the output satisfying specific constraints. Modern deployments increasingly handle the full task in a single LLM call, with one prompt specifying a layered output whose overall artifact, structural sections, and nested fields must each satisfy concrete constraints. Existing instruction-following benchmarks treat the constraint set as a flat list applied uniformly to the response, so they cannot scope a check to a particular section of the output. We introduce IFHierBench, a hierarchical instruction-following benchmark of 600 prompts stratified across four constraint-tree depths and 35 distinct constraints, each prompt paired with a deterministic checker that verifies satisfaction at every scope. Evaluating seven leading proprietary and open-weight models, we find that even the strongest model only marginally exceeds 50% prompt-level accuracy and that accuracy degrades sharply as constraint depth grows. Reliably following nested constraints remains a substantial gap for current LLMs, motivating future training methods that consider constraint adherence at finer granularity to achieve better instruction-following ability.
Following complex instructions with multiple explicit constraints remains a fundamental challenge for large language models (LLMs). Existing alignment methods, such as DPO, optimize holistic reward signals that often underemphasize strict satisfaction of individual constraints, particularly under out-of-distribution or multi-constraint settings. In this paper, we propose STAIF, a stage-wise optimization framework that decouples the alignment of subjective (soft) constraints from the optimization of objectively verifiable (hard) constraints. Stage 1 applies preference optimization with multiple negative samples to sharpen sensitivity to soft constraints, while Stage 2 applies Reinforcement Learning with Verifiable Rewards (RLVR) to enforce strict compliance with hard constraints. To support this method, we construct STAINSTRUCT, a high-quality bilingual (English, Chinese) dataset of approximately 31,000 complex multi-constraint instructions. Extensive analyses validate the design of STAIF and show state-of-the-art performance on representative benchmarks against strong baselines, as well as genuine generalization.
Jian Hong, Chen Cheng, Quan Liu +2
University of Science and Technology of China · iFLYTEK Research Group
Common instruction following (IF) benchmarks, such as IFEval, evaluate the ability of language models (LMs) to follow simple constraints that can easily be verified programmatically. These same constraints, in turn, are used as reward signals for reinforcement learning. This design may lead to reward hacking of constraints, pushing models away from the main task. Recent attempts to address this resulted in benchmarks with richer constraints, but required expert labor, remained closed-access, and cannot be easily extended. We present Lift-Bench, an IF benchmark that asks a model to draft a document given a topic and a structured outline. Constraints are scoped to certain document sections, providing finer control over constraint density. Lift-Bench includes both verifiable and judged constraints, as well as a training counterpart, Lift-Train. Importantly, Lift-Bench provides both main task and instruction-following constraints, mitigating shallow reward hacking. On experiments with multiple models, we find that frontier models struggle with the harder subsets of our benchmark and analyze model performance along multiple axes. We also find that training on Lift-Train is much less susceptible to reward hacking than training on the common IFTrain. Lift-Bench and Lift-Train are publicly available, as well as the data generating pipeline.