Compliance versus Sensibility: On the Reasoning Controllability in Large Language Models
| Authors | Xingwei Tan et al. |
| Year | 2026 |
| HF Upvotes | 5 |
| arXiv | 2604.27251 |
| Download | |
| Code | https://github.com/Xingwei-Tan/compliance_sensibility |
Abstract
Large Language Models (LLMs) are known to acquire reasoning capabilities through shared inference patterns in pre-training data, which are further elicited via Chain-of-Thought (CoT) practices. However, whether fundamental reasoning patterns, such as induction, deduction, and abduction, can be decoupled from specific problem instances remains a critical challenge for model controllability, and for shedding light on reasoning controllability. In this paper, we present the first systematic investigation of this problem through the lens of reasoning conflicts: an explicit tension between parametric and contextual information induced by mandating logical schemata that deviate from those expected for a target task. Our evaluation reveals that LLMs consistently prioritize sensibility over compliance, favoring task-appropriate reasoning patterns despite conflicting instructions. Notably, task accuracy is not strictly determined by sensibility, with models often maintaining high performance even when using conflicting patterns, suggesting a reliance on internalized parametric memory that increases with model size. We further demonstrate that reasoning conflicts are internally detectable, as confidence scores significantly drop during conflicting episodes. Probing experiments confirm that reasoning types are linearly encoded from middle-to-late layers, indicating the potential for activation-level controllability. Leveraging these insights, we steer models towards compliance, increasing instruction following by up to 29%. Overall, our findings establish that while LLM reasoning is anchored to concrete instances, active mechanistic interventions can effectively decouple logical schemata from data, offering a path toward improved controllability, faithfulness, and generalizability.
Engineering Breakdown
Plain English
This paper investigates whether Large Language Models can decouple fundamental reasoning patterns (induction, deduction, abduction) from specific problem instances, which is critical for understanding and controlling how LLMs reason. The authors introduce the concept of 'reasoning conflicts'—situations where they force models to follow logical schemata that contradict what the task actually requires—to systematically study this problem. Their key finding is that LLMs consistently prioritize sensibility (alignment with real-world patterns) over strict compliance with mandated reasoning rules, revealing a fundamental tension between what's encoded in model parameters versus what's provided in context.
Core Technical Contribution
The paper's novel contribution is the first systematic framework for studying reasoning decoupling through reasoning conflicts—an explicit methodology to isolate and measure how LLMs handle contradictions between parametric reasoning knowledge and contextual reasoning instructions. Rather than simply probing whether models can reason, the authors deliberately create misalignment scenarios to expose the underlying mechanisms and hierarchy of reasoning priorities. This reveals that LLMs have implicit preferences for certain reasoning patterns that override explicit instructions, which has direct implications for model controllability and interpretability.
How It Works
The methodology works by constructing tasks where the optimal reasoning pattern differs from the model's expected pattern, creating a controlled conflict. Researchers provide explicit Chain-of-Thought prompts that mandate a specific logical schema (e.g., deductive reasoning) while the parametric knowledge in the model weights 'prefers' a different approach (e.g., inductive reasoning). The model then generates reasoning steps, and researchers measure whether outputs follow the contextual instruction or revert to parametric preferences. By systematically varying reasoning types and measuring compliance rates, they map out a hierarchy showing which reasoning patterns models prefer and how strongly context can override those defaults.
Production Impact
For production systems, this work has immediate implications for prompt engineering and reasoning reliability. If your system relies on steering model behavior through CoT prompts, you now know that models have hard limits—some reasoning patterns can be overridden by prompts while others resist instruction and follow learned patterns instead. This means you cannot blindly trust that carefully crafted reasoning prompts will always work; you need to empirically validate whether the desired reasoning schema is actually controllable for your specific domain. For safety-critical applications (medical reasoning, financial analysis, legal advice), this reveals a controllability gap: models may silently default to learned patterns rather than following your mandated reasoning logic, potentially giving false confidence in steering. The trade-off is that understanding this hierarchy lets you either choose tasks where reasoning patterns align naturally, or invest in fine-tuning if you need guaranteed compliance with non-default reasoning schemas.
Limitations and When Not to Use This
The paper does not address how to actually fix the controllability problem—knowing that models prioritize sensibility over compliance is useful, but the authors don't provide techniques to force adherence to arbitrary reasoning schemas when they conflict with learned patterns. The scope is also limited by the abstract cutoff, but reasoning conflicts are likely most pronounced in out-of-distribution scenarios; the approach may not reveal issues on naturalistic tasks where contextual and parametric reasoning naturally align. Additionally, the paper appears to focus on smaller-scale investigation (typical of foundational work), leaving open questions about scaling: do larger models have different reasoning priorities? Do different training objectives (RLHF, instruction-tuning) change the hierarchy of reasoning preferences? The framework also doesn't address multi-step reasoning or compound conflicts where several reasoning schemas need to work together.
Research Context
This work builds on a decade of research into LLM reasoning capabilities, extending beyond prior probing methods (like BERTology) by moving from passive analysis to active conflict injection. It relates directly to growing concerns about model interpretability and controllability (highlighted by work on mechanistic interpretability and representation theory), and complements recent research on instruction-following and prompt robustness. The paper opens a new research direction: if reasoning patterns form a learnable hierarchy with some more deeply rooted than others, future work can investigate which architectural choices or training regimes alter this hierarchy, or whether it's a fundamental property of transformer-based language models. This also connects to broader safety research—understanding reasoning controllability is essential for alignment and verifiability of model outputs in deployment.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::