Research map / Large language models / Reasoning and chain of thought

Trade-offs in Large Reasoning Models: An Empirical Analysis of Deliberative and Adaptive Reasoning over Foundational Capabilities

AAAI 2026 · Large language models › Reasoning and chain of thought · approach: Long CoT dynamics and efficiency · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Empirically evaluates how distillation and RL-based deliberative reasoning affect helpfulness and safety of 7B-70B reasoning models, and shows that adaptive modes (Zero-, Less-, Summary-Thinking) recover some lost performance.

Problem

Deliberative long chain-of-thought reasoning in large reasoning models may degrade foundational capabilities (helpfulness, harmlessness) and increase inference cost on general tasks, a trade-off that is under-studied.

Approach

Benchmarks LRMs against their chat-model counterparts on general, instruction-following, and safety tasks, then manipulates the thinking segment at inference (inserting </think> early, summarizing thoughts, or keeping the first sentence) to show task-dependent adaptive reasoning modes can improve results without retraining.

Key results

• R1-Distill-Llama-70B's WildJailbreak score rises from 28.25 to 89.10 under Zero-Thinking, versus 19.5 for Llama-3.3-70B-Instruct.
• Summary-Thinking improves s1.1-32B IFEval performance by 59.74% relative, and QwQ-32B by 4.44%.
• s1.1-32B shows a 47.38% decrease in IFEval versus its chat base Qwen2.5-32B-Instruct, with about 250% higher inference cost.

Surprising findings

• Zero-Thinking greatly boosts safety rejection rates while reducing helpfulness on general tasks.
• Thoughts are more often unsafe than final responses, with unsafe thoughts leading to unsafe answers.

Limitations

• Evaluation limited to general tasks, instruction following, safety, and reasoning; excludes multimodal and open-ended dialogue.
• Results limited to specific checkpoints and do not cover MoE or multimodal LRMs.
• Uses GPT-4o as judge for MT-Bench and harmlessness evaluation.

Experimental setup

Datasets: MMLU-Pro LiveBench IFEval MT-Bench StrongReject WildJailbreak

Benchmarks: MMLU-Pro LiveBench IFEval MT-Bench StrongReject WildJailbreak

Metrics: accuracy prompt-level accuracy MT-Bench score (1-10) rejection rate response and thought length

Baselines: Qwen2.5-7B-Instruct Qwen2.5-32B-Instruct Llama-3.3-70B-Instruct

Builds on: DeepSeek-R1 s1 OpenThinker QwQ-32B LiveBench

Closest papers by meaning

• Think or Not? Exploring Thinking Efficiency in Large Reasoning Models via an Information-Theoretic Lens (NeurIPS 2025 · similarity 0.92)
• The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity (NeurIPS 2025 · similarity 0.91)
• Don’t Think Longer, Think Wisely: Optimizing Thinking Dynamics for Large Reasoning Models (NeurIPS 2025 · similarity 0.91)
• On Reasoning Strength Planning in Large Reasoning Models (NeurIPS 2025 · similarity 0.91)
• Measuring the Faithfulness of Thinking Drafts in Large Reasoning Models (NeurIPS 2025 · similarity 0.91)
• Training Language Models to Reason Efficiently (NeurIPS 2025 · similarity 0.90)