Skip to main content
Back

Tencent's CRGC Cuts AI Constraint Violations by 39% Without Retraining

A multi-institution research team introduces CRGC, a graph-based prompting framework that cuts LRM constraint violations by 39% without any model retraining

Paper, Full Access
  • Researchers from CUHK, Tencent Jarvis Lab, Westlake University, and King's College London introduce CRGC, a graph-based framework for fixing multi-constraint instruction following in LRMs.
  • The paper formalizes the Constraint Adherence Problem (CAP): models either ignore individual constraints or fail to balance conflicting ones simultaneously.
  • CRGC decomposes prompts into a directed constraint graph, detects problematic edges, and auto-generates natural-language "bridge constraints" to reconcile conflicts before generation.
  • Results: 39% fewer constraint violations vs. standard prompting, 27.7% fewer vs. Chain-of-Thought, with no degradation on general reasoning benchmarks.
  • Compared to iterative repair methods (DeCRIM), CRGC resolves conflicting constraints 19.7% better and requires 1.2 inference turns on average vs. 4.2, cutting total token use by 67%.
  • No model retraining required; works with any API-accessible model. Code is open-sourced on GitHub.

Large Reasoning Models (LRMs) like GPT-4o, Claude 3.7, and DeepSeek-R1 can solve graduate-level math and write code from scratch. Ask them to do something with five specific formatting rules at once, though, and they start quietly ignoring some of them. A research team from the Chinese University of Hong Kong, Tencent Jarvis Lab, Westlake University, and King's College London has published a paper formalizing this failure mode and proposing a structured fix that requires zero retraining.

The problem has a name now

The researchers call it the Constraint Adherence Problem (CAP). It shows up in two distinct ways: a model either overlooks a constraint entirely (say, forgetting to enclose the answer in double quotes), or it fails to balance two constraints that seem to pull in opposite directions (like being both "concise" and "detailed"). This is not a niche edge case. Evaluations of 23 recent LRMs have uncovered a widespread inability to follow user constraints, especially on harder problems and multi-constraint settings. A separate benchmark found that across open-source LRMs including GPT-OSS, Qwen3, and DeepSeek-R1, the highest instruction following score remains below 0.25, meaning fewer than 25% of reasoning traces comply with given instructions.

The deeper tension is structural. Improving reasoning capability often comes at the cost of instruction adherence, suggesting an inherent trade-off between the two abilities. Training for stronger reasoning alone may undermine alignment, and future methods must account for this tension to build models that are both capable and controllable. Fine-tuning on instruction-following data is the obvious fix, but it reliably degrades general reasoning performance by 6-7% in the paper's own comparisons.

A constraint graph, not a bigger prompt

The team's solution is the Constraint Relationship Graph Completion (CRGC) framework. The core idea: before sending a prompt to the model, decompose it into individual constraints, build a directed graph that maps how each constraint influences the others, and then automatically generate short "bridge constraints" to patch the problematic edges. The whole thing happens at inference time, with no weight updates.