Concise and Adaptive Thinking

• Concise and adaptive thinking is a reasoning approach that adjusts inference steps based on query difficulty, optimizing token cost and performance.
• It employs control signals and adaptive thresholds to selectively switch between fast, implicit responses and slower, explicit chain-of-thought reasoning.
• Recent research shows that this strategy reduces inference tokens and latency while improving accuracy, addressing the overthinking challenge in large models.

Searching arXiv for papers on concise and adaptive reasoning in large language and reasoning models. Searching arXiv by topic and specific identifiers to verify the cited works. Concise and adaptive thinking denotes a class of reasoning strategies in which a model allocates reasoning effort according to input characteristics such as difficulty, uncertainty, task structure, or interaction context, rather than applying a uniform long Chain-of-Thought (CoT) to every query. In the recent literature, this agenda is motivated by the overthinking problem: large reasoning models often generate unnecessarily long traces for simple problems, increasing latency and token cost, while still lacking mechanisms for selectively deepening reasoning on harder instances. Surveys now frame the area as a shift from pure efficiency—shortening reasoning traces in the abstract—to adaptivity, namely input-conditioned control over whether, when, and how much to think (Wu et al., 13 Nov 2025, Zhu et al., 13 Jul 2025).

1. Conceptual foundations

A standard distinction in this literature is between “fast-thinking” and “slow-thinking.” Fast-thinking refers to direct, implicit answer generation without explicit intermediate steps (“No-CoT”), with low inference cost and minimal latency. Slow-thinking invokes an explicit CoT before the final answer, increasing compute but often improving performance on difficult reasoning tasks (Zhu et al., 13 Jul 2025). Many methods can be interpreted as learning a policy over this continuum, rather than treating one mode as globally optimal.

The underlying trade-off is often written as an input-specific optimal reasoning length,

$L^*(x)=\arg\max_L \bigl[A(L\mid x)-\beta L\bigr],$

where accuracy and reasoning cost are jointly optimized (Zhu et al., 13 Jul 2025). This formulation reflects a core empirical observation repeated across studies: simple questions are often harmed by unnecessary reflection, while hard questions require additional deliberation, verification, or search.

Several papers ground this distinction in cognitive structure. Adaptive Mode Learning (AML) defines four hierarchical thinking modes inspired by Hierarchical Cognitive Control Theory: Mode 1 “Intuitive Response,” Mode 2 “Intentional Analysis,” Mode 3 “Strategic Adaptation,” and Mode 4 “Prospective Simulation.” These modes are aligned respectively with sensory–motor, contextual, episodic, and branching control in the account attributed to Koechlin & Summerfield and Badre, and each mode has a predefined sequence of thinking steps and think-tags for explicit token accounting (Wang et al., 4 May 2025). In a different cognitive lineage, CogAct models adaptive conceptual processing through chunking, attention, STM, and LTM, emphasizing chunk formation economy, hierarchical knowledge, sliding attention, and subjectivity by design (Bennett et al., 21 Dec 2025). This suggests that current machine-learning approaches and older cognitive architectures converge on a shared principle: reasoning should be selective, structured, and resource-bounded rather than uniformly exhaustive.

2. Formal objectives and control signals

The recent survey literature formalizes adaptive reasoning as a control-augmented policy optimization problem:

$$\max_{\phi\in\Phi}\; \mathbb{E}_{x\sim\mathcal D,\;r\sim\pi_\theta(\cdot\mid x;\phi(x))} \bigl[\mathcal P(r,x)-\lambda\,\mathcal C(r,x)\bigr],$$

where $\mathcal P(r,x)$ measures task performance and $\mathcal C(r,x)$ measures compute cost (Wu et al., 13 Nov 2025). A related survey gives the token-cost decomposition

$$\mathrm{Cost}(r,y)=\lambda_{\rm cot}L+\lambda_{\rm ans}T,$$

separating reasoning-token cost from answer-token cost (Zhu et al., 13 Jul 2025). The central design choice in most systems is therefore not whether to optimize for correctness or efficiency, but which control signal should mediate the trade-off.

One family of signals is outcome-based. “Correct, Concise and Complete” uses a reward

$R(y)=R_C(y)-\lambda R_L(y),$

where $R_C$ verifies correctness and format, and $R_L$ penalizes tokens generated after the first correct answer appears (Rakotonirina et al., 6 Jan 2026). This reward targets overthinking after success rather than reasoning length in the abstract. “Thinking Fast and Right” replaces a static penalty with an adaptive coefficient $\lambda_t$ that increases when batch accuracy exceeds a reference level and relaxes when accuracy drops, thereby reshaping the accuracy–length trade-off during training (Su et al., 23 May 2025).

A second family is sufficiency-based. SuCo defines the Minimal Sufficient CoT (MSC) as the shortest prefix whose sufficiency score is above a threshold $\delta$, and makes that threshold input-adaptive:

$$\max_{\phi\in\Phi}\; \mathbb{E}_{x\sim\mathcal D,\;r\sim\pi_\theta(\cdot\mid x;\phi(x))} \bigl[\mathcal P(r,x)-\lambda\,\mathcal C(r,x)\bigr],$$0

Difficulty is estimated by a length percentile over full-CoT traces, so easy questions receive lower sufficiency thresholds and hard questions higher ones (Wang et al., 16 Jun 2026). This replaces a fixed budget with a criterion for when reasoning is sufficient.

A third family is stepwise process control. SAT defines a step difficulty score as $$\max_{\phi\in\Phi}\; \mathbb{E}_{x\sim\mathcal D,\;r\sim\pi_\theta(\cdot\mid x;\phi(x))} \bigl[\mathcal P(r,x)-\lambda\,\mathcal C(r,x)\bigr],$$1, where $$\max_{\phi\in\Phi}\; \mathbb{E}_{x\sim\mathcal D,\;r\sim\pi_\theta(\cdot\mid x;\phi(x))} \bigl[\mathcal P(r,x)-\lambda\,\mathcal C(r,x)\bigr],$$2 is a lightweight Process Reward Model estimate that a step leads to a correct final answer; that score then drives transitions among FAST, NORMAL, SLOW, and SKIP states in a finite-state controller (Huang et al., 9 Apr 2026). In these formulations, adaptive reasoning becomes a control problem over trajectories rather than a one-shot choice between short and long generation.

3. Discrete mode selection and routing mechanisms

Many systems implement adaptivity through a small set of explicit reasoning modes. These approaches differ primarily in what the modes represent—thinking depth, response format, or latent strategy—and how switching is learned.

Family	Mechanism	Representative papers
Hierarchical thinking modes	Discrete modes from intuitive response to deep deliberation	AML (Wang et al., 4 May 2025)
Binary think / no-think control	Choose between explicit CoT and direct answer	AdaptThink (Zhang et al., 19 May 2025), AutoThink (Tu et al., 16 May 2025)
Format selection	Choose Full Format, Perception-Only Format, or Direct Answer	AVR (Huang et al., 16 Apr 2026)
Fast / slow routing	Route between codebook-guided fast inference and explicit reasoning	“Fast Thinking for LLMs” (Zheng et al., 28 Sep 2025)

AML is the most explicit depth-stratified version. Its Adaptive Mode Policy Optimization (AMPO) decomposes advantage into sample-level and mode-level terms, allowing the policy to learn context-aware mode switching while penalizing unnecessary chain length when modes tie on reward (Wang et al., 4 May 2025). Empirically, early dialogue turns favor Mode 4 over 50% of the time, whereas later turns see Mode 1 chosen nearly half of the time; hard social scenarios prefer deep modes, while “both sides win” scenarios are dominated by shallow modes.

AdaptThink reduces the decision to a binary action at the first generated token: $$\max_{\phi\in\Phi}\; \mathbb{E}_{x\sim\mathcal D,\;r\sim\pi_\theta(\cdot\mid x;\phi(x))} \bigl[\mathcal P(r,x)-\lambda\,\mathcal C(r,x)\bigr],$$3 indicates NoThinking, whereas a start token such as “Alright” indicates Thinking. Its constrained optimization objective encourages NoThinking while maintaining reference-level performance, and its importance-sampling policy enforces 50% Thinking and 50% NoThinking at cold start so that both modes remain explorable during on-policy RL (Zhang et al., 19 May 2025). AutoThink finds a related binary gate in R1-style models by replacing the usual CoT seed with an ellipsis inside > ..., which produces a stochastic switch between no-thinking and thinking modes; RL then sculpts this initially uninformed gate into a difficulty-aware policy (Tu et al., 16 May 2025).

In multimodal reasoning, AVR decomposes visual reasoning into visual perception, logical reasoning, and answer application, then lets the model choose among Full Format, Perception-Only Format, and Direct Answer (Huang et al., 16 Apr 2026). “Fast Thinking for LLMs” takes a different route: it learns a Latent Codebook of strategy priors from concise rationale “sketches,” injects continuous thinking vectors into the hidden state, and uses GainRouter to decide whether the fast codebook path is sufficient or whether explicit slow CoT is needed (Zheng et al., 28 Sep 2025). Discrete mode policies therefore span both symbolic response formats and hidden-state routing architectures.

4. Continuous sufficiency, pruning, and compression

Other systems reject discrete tiers and instead learn continuous or prefix-level control over reasoning depth. SuCo is the clearest statement of this position. Its two-stage pipeline first constructs MSC-aligned data through MSC-Aligned Fine-Tuning and then applies Sufficiency-Aware Policy Optimization, whose reward penalizes both over-thinking beyond a tolerance and under-thinking when the answer is wrong (Wang et al., 16 Jun 2026). The paper argues that discrete modes or fixed budget tiers lack a principled sufficiency criterion.

“Think When You Need” formalizes length control as pairwise comparison rather than direct penalty. In the verifiable setting, rewards compare sampled explanations on correctness and, among correct ones, on length; in fuzzy tasks without ground truth, the same framework replaces correctness with pairwise quality judgments or a Bradley–Terry reward model (Yang et al., 4 Apr 2025). The important point is that verbosity is never penalized independently of solution quality.

ConCISE targets a different mechanism: redundant reflection emerging from “Confidence Deficit” and “Termination Delay.” It injects confidence phrases when internal stepwise confidence falls below a threshold and halts when a probing detector indicates sufficient answer confidence, with $$\max_{\phi\in\Phi}\; \mathbb{E}_{x\sim\mathcal D,\;r\sim\pi_\theta(\cdot\mid x;\phi(x))} \bigl[\mathcal P(r,x)-\lambda\,\mathcal C(r,x)\bigr],$$4 in the paper’s early-stopping criterion (Qiao et al., 8 May 2025). “Don’t Think Longer, Think Wisely” segments reasoning traces into modular “thinking patterns,” estimates candidate exit points by Monte Carlo continuation, appends finalize patterns, and then prunes detrimental segments through an auxiliary LLM that labels segments as KEEP or REMOVE (An et al., 27 May 2025). These methods compress reasoning not by selecting one of a few predefined modes, but by learning when a partial trajectory has already become sufficient or by deleting locally unhelpful substructures.

Draft-Thinking occupies an intermediate position. It first teaches a “draft-style” sparse reasoning pattern by supervised fine-tuning, then restores capability through a progressive RL curriculum with increasing $$\max_{\phi\in\Phi}\; \mathbb{E}_{x\sim\mathcal D,\;r\sim\pi_\theta(\cdot\mid x;\phi(x))} \bigl[\mathcal P(r,x)-\lambda\,\mathcal C(r,x)\bigr],$$5, and finally uses an adaptive prompt instructing the model to decide whether to use draft or full step-by-step reasoning based on question difficulty (Cao et al., 28 Feb 2026). The architecture is unchanged; what changes is the learned default granularity of reasoning.

5. Empirical regularities across domains

A striking regularity across the literature is that adaptive control often improves both efficiency and task performance, rather than merely trading one for the other. In social interaction, AML reports 15.6% higher task performance than GPT-4o on the Goal metric over GPT-4o in self-play on SOTOPIA-Hard, and AMPO improves Goal by 7.0% over GRPO while using 67.2% as many inference tokens (581 vs 865), a 32.8% reduction in chain length (Wang et al., 4 May 2025).

In mathematical reasoning, several binary or multi-stage methods report large token savings with stable or improved accuracy. AdaptThink reduces the average response length of DeepSeek-R1-Distill-Qwen-1.5B by 53% and improves its accuracy by 2.4% across GSM8K, MATH500, and AIME2024 (Zhang et al., 19 May 2025). AutoThink reports a 6.4 percent relative accuracy improvement while reducing token usage by 52 percent on DeepSeek-R1-Distill-Qwen-1.5B (Tu et al., 16 May 2025). ConciseR moves from a base average score of approximately 37.5% to 56.1% after its two RL stages, while Stage 2 reduces CoT length by 20–23% relative to Stage 1 across AIME 24, MATH-500, AMC 23, Minerva, and OlympiadBench (Song et al., 27 May 2025). “Correct, Concise and Complete” reports average response-length reductions of 28% for Qwen3-8B and 40% for Qwen3-32B, with accuracy drops of 1.6 and 2.5 points respectively, and an $$\max_{\phi\in\Phi}\; \mathbb{E}_{x\sim\mathcal D,\;r\sim\pi_\theta(\cdot\mid x;\phi(x))} \bigl[\mathcal P(r,x)-\lambda\,\mathcal C(r,x)\bigr],$$6 of 76.6 for its best variant (Rakotonirina et al., 6 Jan 2026).

SuCo extends these gains beyond a single benchmark family. On GSM8K at 1.5B scale it reaches 87.7% using 304 tokens versus 501 for full-CoT distillation; on AIME25 at 7B it reaches 61.7% using 2,679 tokens versus 10,833; on MBPP it reports 48.5% versus 41.0% with 930 tokens versus 3,511; on MMLU-STEM it reports 38.6% versus 33.8% with 745 tokens versus 1,956 (Wang et al., 16 Jun 2026). The same paper reports that truncating to MSC prefixes cuts reasoning tokens by 60–80% at each difficulty level while improving accuracy by 2–5 points, and that trivial queries often omit intermediate steps entirely for approximately 30–37% of samples.

Adaptive reasoning is not confined to text-only math. AVR reports 50–90% token reduction across OCRBench, TextVQA, MathVista, MathVision, MMMU, MMMU-Pro, and MMBench; for Qwen3-VL-4B it improves OCRBench from 81.3% at 476.9 tokens to 85.6% at 59.4 tokens and MMMU from 58.3% at 1332.1 tokens to 59.8% at 128.7 tokens (Huang et al., 16 Apr 2026). In egocentric video understanding, EgoReasoner’s task-adaptive templates and task-aware rewards yield 37.5% average accuracy on HD-EPIC, exceeding Qwen2.5-VL-7B at 25.7% while using a 3B model trained on 16K samples (Zhu et al., 6 Mar 2026). A plausible implication is that concise and adaptive thinking is better understood as a general control principle for structured inference than as a narrow optimization for text-only CoT.

6. Failure modes, evaluation, and open problems

A recurrent theme in the literature is that naive length reduction is not equivalent to adaptive reasoning. Fixed penalties, fixed thresholds, and fixed mode schedules are repeatedly shown to be brittle. SuCo reports that static thresholds $$\max_{\phi\in\Phi}\; \mathbb{E}_{x\sim\mathcal D,\;r\sim\pi_\theta(\cdot\mid x;\phi(x))} \bigl[\mathcal P(r,x)-\lambda\,\mathcal C(r,x)\bigr],$$7 are rigid: a high $$\max_{\phi\in\Phi}\; \mathbb{E}_{x\sim\mathcal D,\;r\sim\pi_\theta(\cdot\mid x;\phi(x))} \bigl[\mathcal P(r,x)-\lambda\,\mathcal C(r,x)\bigr],$$8 wastes tokens on easy questions, while a low $$\max_{\phi\in\Phi}\; \mathbb{E}_{x\sim\mathcal D,\;r\sim\pi_\theta(\cdot\mid x;\phi(x))} \bigl[\mathcal P(r,x)-\lambda\,\mathcal C(r,x)\bigr],$$9 under-thinks on hard ones (Wang et al., 16 Jun 2026). “Thinking Fast and Right” states that static direct length penalties are hard to tune and can cause sharp collapse in both length and accuracy after approximately 100 steps, whereas its adaptive penalty converges more smoothly (Su et al., 23 May 2025). ConciseR explicitly notes that directly rewarding short length led to “empty CoT” cheating (Song et al., 27 May 2025).

Mode-based systems have analogous failure modes. AdaptThink reports that without importance sampling the model never sees NoThinking initially and collapses to pure Thinking (Zhang et al., 19 May 2025). AutoThink shows that without Stage 1 reward balancing, the policy collapses into a single mode, and bypassing Stage 2 reduces accuracy from approximately 51.7% to approximately 47.6% (Tu et al., 16 May 2025). SAT adds that static thresholds remain a limitation, and its external Pilot introduces minor inference overhead while relying on the base model’s instruction-following for control tags (Huang et al., 9 Apr 2026).

Evaluation methodology has therefore expanded beyond plain accuracy and mean token count. The survey literature lists pass@1, inference tokens, latency, speed-up, CoT Trigger Rate, and several concise-accuracy metrics (Zhu et al., 13 Jul 2025). “Correct, Concise and Complete” introduces Overthinking-Adjusted Accuracy and $\mathcal P(r,x)$0 to assess the joint quality–efficiency frontier (Rakotonirina et al., 6 Jan 2026), while AutoThink defines Efficiency-F1 to summarize simultaneous gains in accuracy and token use (Tu et al., 16 May 2025). This suggests that adaptive reasoning requires evaluation at the level of policies and trajectories, not just final answers.

Two recent surveys locate the field’s main open problems in self-evaluation, meta-reasoning, and human-aligned reasoning control. The reported issues include miscalibrated intermediate confidence, weak internal error detection, the unresolved problem of deciding not only how long to think but which reasoning paradigm to invoke, and the need to preserve trust and transparency when adaptive truncation omits steps (Wu et al., 13 Nov 2025, Zhu et al., 13 Jul 2025). In that sense, concise and adaptive thinking is no longer only a compression problem. It is increasingly treated as a control problem over reasoning form, depth, and termination, with correctness, cost, and interpretability jointly in view.