Source-linked AI summary
Compliance versus Sensibility: On the Reasoning Controllability in Large Language Models
Xingwei Tan, Marco Valentino, Mahmud Elahi Akhter, Yuxiang Zhou, Maria Liakata, Nikolaos Aletras
TL;DR
The paper asks whether deduction, induction, and abduction can be controlled independently of specific problem instances. It evaluates instruction–task conflicts, probes their internal representations, and finds that steering can increase instruction following by up to 29%, although with generation-quality costs.
Problem
It remains unclear whether CoT reasoning patterns are disentangled capabilities or intrinsically tied to the tasks and instances seen during pre-training, limiting explicit structural controllability.
Method
The authors evaluate aligned and conflicting reasoning instructions on logical tasks, compare compliance with sensibility, probe middle-to-late representations, and apply Contrastive Activation Addition.
Results
Up to 29% higher reasoning-type compliance is achieved through steering, while models generally prioritize task-appropriate sensibility over conflicting instructions.
Takeaways & Limitations
Reasoning conflicts are active and mitigable through activation steering, supporting a route toward more controllable reasoning engines.
Takeaways & Limitations
Contrastive Activation Addition can degrade generation quality, including formatting and task neglect, in the localized open-weight architectures primarily tested.
Abstract
from arXiv · showhide
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. 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.
1 Introduction
The paper studies whether LLMs can follow mandated reasoning types when those instructions conflict with task-appropriate logic. It evaluates this conflict systematically and investigates mechanistic representations and activation steering for improving compliance.
- The study asks whether CoT reasoning patterns are generalizable and controllable or remain tied to specific tasks and instances.This matters because standard CoT leaves the model to choose deduction, induction, or abduction, raising robustness and faithfulness concerns.
- The authors construct logical questions requiring deduction, induction, or abduction, then compare responses under aligned and conflicting reasoning instructions.The testbed evaluates both sensibility, meaning task-appropriate reasoning, and compliance, meaning following the instruction.
- The evaluation covers three open-source model families, two frontier LLMs, and four logical reasoning datasets.This provides a broad assessment of how models direct reasoning across diverse logical questions.
- Probing examines how reasoning instructions and internal reasoning priors interact, while quantifying compliance across reasoning types and instruction settings.The paper investigates whether models encode these instructions and how they decide between compliance and task-appropriate reasoning.
- Activation steering increases reasoning-type compliance by up to 29%.The paper uses Contrastive Activation Addition to augment compliance during inference.
2 Related Work
Prior work identifies limitations in logical reasoning and CoT faithfulness, while probing and model-editing research motivates studying reasoning representations and steering.
- LLMs often struggle with logical and soft reasoning, including identifying the correct reasoning direction, and no single strategy fits every context.Prior studies report a substantial performance gap across fundamental reasoning paradigms, especially abduction.
- CoT rationales may function as post-hoc justifications rather than faithful descriptions of the internal reasoning process.Other work nevertheless suggests CoT can actively guide outputs even without a direct causal link.
- Probing studies place instruction-following and task-specific logic primarily in residual-stream activations that mature in middle-to-late transformer layers.This motivates connecting representation analysis with controllability methods.
- Model editing offers a lightweight alternative to fine-tuning, and this paper applies Contrastive Activation Addition to the abstract domain of reasoning compliance.Earlier steering work often targeted binary attributes or content effects rather than reasoning compliance.
3 Methodology
The methodology measures reasoning controllability by comparing mandated and task-suitable reasoning types, then analyzes and steers compliance under conflicts.
- 3.1 Problem Formulation: The study evaluates whether models follow mandated reasoning types when those instructions may conflict with the task-appropriate type.Questions are associated with a suitable type t, while prompts can mandate a different type t′.
- 3.3 Automated Reasoning Type Classification: Responses are categorized by whether their generated reasoning matches the mandated type, the suitable type, both, or neither.These categories distinguish sensible and compliant behavior, including cases where the model follows an ill-suited instruction.
- 3.2 Prompting for Reasoning Controllability: Zero-shot prompts provide high-level definitions and general procedures for deductive, inductive, and abductive reasoning.The prompts are used to test in-context reasoning controllability.
- 3.3 Automated Reasoning Type Classification: Generated CoT is classified as deductive, inductive, abductive, or direct using an LLM judge and a rule-based output parser.The judge receives the generated reasoning and predicts its reasoning type.
- 3.4 Steering for Reasoning Controllability: CAA steers compliance by deriving a mean-difference activation vector from contrastive compliant and non-compliant responses and adding it during inference.A coefficient controls the steering effect, while pairing controls for answer correctness and instructed reasoning type.
4 Experimental Setup
The experiments compare frontier and open-weight models across four datasets representing deduction, induction, and abduction, using accuracy, sensibility, compliance, confidence, and probing measures.
- Models: The evaluation includes GPT-5.1, GEMINI3-FLASH, and OLMO, QWEN, and LLAMA families across multiple model sizes and modes.Open-weight comparisons include smaller and larger models, instruct and thinking versions, and QWEN with thinking on or off.
- Datasets: Four datasets represent deduction, induction, abduction, and mixed deduction–abduction reasoning: FOLIO, SPR, αNLI, and RECV.FOLIO tests conclusions from logical premises, SPR tests latent sequence rules, αNLI tests explanatory hypotheses, and RECV covers claim verification.
- Evaluation Metrics: Task accuracy is computed by parsing tagged final answers and comparing them with gold answers.The parser extracts the final answer from generated outputs using the prompt’s special tags.
- Evaluation Metrics: Sensibility and compliance rates are obtained by classifying generated CoT and comparing its inferred reasoning type with the suitable and instructed types.GPT-5.1 judges models other than itself, while GEMINI3-FLASH judges GPT-5.1.
- Evaluation Metrics: Confidence is measured from the probability of generating token A when the model is asked whether its answer is true or false.This probes internal effects of reasoning conflicts beyond the generated textual CoT.
- Probing for Reasoning: Linear probing uses frozen residual-stream representations from every layer, mean-pooled over question-span tokens, to examine encoded reasoning instructions.The representations are collected after the feed-forward block for responses generated under reasoning instructions.
5 Results
Across models and datasets, reasoning sensibility generally outweighs instruction compliance and more strongly predicts accuracy. Confidence and hidden-state analyses reveal detectable but entangled compliance signals, while activation steering can increase compliance.
- Reasoning conflicts: 43.5% of cases matched task-required reasoning without following instructions, versus 18.6% that followed instructions at the expense of task appropriateness.Stray reasoning occurred in 5.7% of samples.
- Reasoning sensibility: Reasoning sensibility generally increased with model size, but architecture and instruction-tuning choices produced notable exceptions.GPT-5.1 reached 88.1% sensibility, while LLAMA3.1-8B-IT reached 58.2%.
- Instruction compliance: Compliance varied substantially across model families and did not reliably increase with parameter count.LLAMA3.1-8B-IT achieved 65.1% compliance, whereas OLMO3-7B-IT achieved 39%.
- Instruction compliance: Built-in thinking generally increased sensibility while reducing compliance, distinguishing it from externally prompted Chain-of-Thought.This pattern appeared in QWEN models and the larger thinking OLMO model.
- Accuracy: Logical sensibility was the primary determinant of accuracy, with sensible reasoning outperforming compliant but unsuitable reasoning across cohorts.On FOLIO, sensible CoT held a 20%–40% accuracy lead over compliant, non-sensible reasoning.
- Confidence: Compliant responses had significantly higher p(True) confidence than non-compliant responses regardless of correctness, with small effect sizes.Cohen’s d was 0.09 for correct responses and 0.12 for incorrect responses.
- Mechanistic analysis: Compliance was moderately decodable from hidden states, typically peaking in middle and late layers, whereas reasoning type was more linearly accessible.The findings support middle-layer intervention while characterizing compliance as weaker and more entangled.
- Mechanistic analysis: Steering layers 14–17 of OLMO3-7B-IT increased compliance with positive µ, while negative µ consistently improved accuracy and encouraged direct answers.For inductive and deductive instructions on α-NLI, accuracy first declined and then peaked as compliance increased.
6 Conclusion
The paper characterizes reasoning conflicts as an active, model-dependent tension between encoded reasoning instructions and compliance. Activation steering mitigates this tension, increasing instruction compliance by up to 29%.
- Models encode reasoning instructions in middle-to-late layers but often struggle to prioritize compliance.
- Reasoning conflicts vary significantly by model scale and architecture and operate as an active mid-computation process.
- Activation steering mitigates reasoning conflicts rather than treating them as immutable limitations.
- 29%: activation steering enhances instruction compliance by up to 29%.
- The study identifies extending the analysis to questions involving multiple reasoning types as future work.
Limitations
The steering intervention improves compliance but has a visible generation-quality cost. Validation is concentrated on localized open-weight architectures, leaving broader stability and scalability questions open.
- Activation steering can improve compliance while visibly reducing generation quality.
- The intervention experiments primarily validate localized open-weight architectures.
- Strong activation steering frequently induces formatting degradation or task neglect in these architectures.
- Future work must test whether scaling laws or dynamic, layer-specific vectors can preserve coherence while isolating reasoning compliance.
A Implementation Details
The implementation uses specified inference hardware, sampling settings, model-steering and probing tools, dataset splits, prompts, and a confidence-estimation procedure.
- Responses from open LLMs are generated with vLLM on NVIDIA GH200, A100, and AMD MI300X hardware.
- Inference temperature is 0.5, and responses are sampled with three random seeds.
- EasyEdit 2.0 is used for model steering, while probing uses logistic regression with C = 1.0 and an L2 penalty.
- SPR uses 4,300 training samples and 100 test samples; RECV has 1,302 training and 99 test samples, while αNLI contributes 100 sampled test questions.
- The evaluation uses complete conflict-inducing prompts and a p(True) follow-up prompt for confidence estimation.
D LLM Judge Agreement
The paper validates automated reasoning-type judgments against human annotations and reports agreement, while the accompanying tables organize judge accuracy and reasoning-conditioned results.
- Figure 7 presents a confusion matrix between aggregated human labels and LLM labels, with human predictions as rows.
- A balanced validation set contains 360 responses sampled across three model families and three fundamental reasoning types.
- Two humans annotate the full validation set, and a third reconciles disagreements to produce the gold standard.
- GEMINI-3-FLASH and GPT-5.1 achieve Cohen’s kappa scores of 0.83 and 0.84 against human annotations.
- Tables 5 and 6 report judge-inferred category accuracy and final-answer accuracy grouped by instructed reasoning type.
F Probing Results
Probing shows that models represent instructed reasoning types clearly even when their final reasoning is non-compliant, revealing a gap between instruction encoding and reasoning behavior.
- Instruction Encoding: In non-compliant cases, instructed reasoning type remains almost perfectly decodable from the residual stream across open model families.This indicates that compliance failures are not primarily caused by failure to represent the instruction.
- Instruction Encoding: A larger gap between instructed-reasoning and judge-inferred reasoning signals indicates clear instruction representation without full realization in final reasoning behavior.The gap is largest in LLaMA, smaller in OLMO, and Qwen is most aligned overall.
- Reasoning Compliance: Judge-inferred reasoning is a noisier proxy for internal computation than the explicit instruction signal, especially in non-compliant cases.The weaker judge-side probability on non-compliant cases supports this distinction.
G Steering Results
The steering experiments evaluate multiplier effects across reasoning types and model–dataset settings, using layer-wise reasoning-type accuracy as part of the analysis.
- Steering Experiments: Steering results for OLMO3-7B-IT on FOLIO examine the impact of multiplier µ across reasoning types.The corresponding experiment is shown in Figure 10.
- Steering Experiments: Steering results for LLAMA3.1-8B-IT on α-NLI examine the impact of multiplier µ across reasoning types.The corresponding experiment is shown in Figure 11.
- Layer-wise Analysis: Layer-wise accuracy results report instructed and judge-inferred reasoning types for each model family.These results are presented in Figure 9.