Adaptive Reasoning Suppression

• Adaptive Reasoning Suppression (ARS) is a family of methods that reduce unnecessary, redundant, or harmful reasoning in large language models to improve efficiency.
• ARS techniques include pre-generation routing, decoding-time suppression, and latent-space control, with methods achieving up to 53% token reduction and 57.9% energy savings.
• The approach balances minimal reasoning for simple tasks with deeper computation for complex queries, ensuring accuracy while significantly lowering compute costs.

Searching arXiv for the cited works to ground the article in current papers. First, I’ll verify the main ARS paper and then gather closely related adaptive reasoning-control papers. Adaptive Reasoning Suppression (ARS) denotes methods that reduce unnecessary, redundant, or harmful reasoning in large reasoning models without eliminating the reasoning needed for difficult inputs. In the narrow sense, the namesake method “ARS: Adaptive Reasoning Suppression for Efficient Large Reasoning LLMs” is a training-free decoding procedure that monitors certainty at multiple checkpoints and suppresses reflection-trigger tokens when certainty becomes high enough (Zheng, 29 Sep 2025). In the broader literature, closely related methods pursue the same objective through pre-generation format choice, adaptive length-aware reinforcement learning, early exit, latent-state steering, and reasoning-trace unlearning (Wu et al., 26 May 2025, Zhang et al., 21 May 2025, Li et al., 13 Jun 2026, Wang et al., 15 Jun 2025). A terminological complication is that “ARS” also names “Answer-agreement Representation Shaping,” a hallucination-detection method rather than a reasoning-suppression method (Zhang et al., 24 Jan 2026).

1. Overthinking, selective compute, and the scope of suppression

ARS research is organized around the overthinking problem in large reasoning models. These models often generate long chains of thought with reflection, backtracking, and self-verification even when the task is simple, which increases token cost, latency, and energy use; several papers also report that unnecessary extra reasoning can degrade accuracy on easy inputs (Zheng, 29 Sep 2025). Related work repeatedly distinguishes the desired regime from both under-reasoning and blanket truncation: hard tasks still benefit from extended reasoning, whereas easy tasks often do not (Wu et al., 26 May 2025).

A recurring conceptual distinction is between suppressing visible reasoning and suppressing reasoning computation itself. Several ARS-relevant papers act primarily on visible generated traces. ARM, for example, does not halt an already-started chain of thought; it teaches the decoder to choose among Direct Answer, Short CoT, Code, and Long CoT before generation, so suppression occurs by avoiding an expensive format from the outset (Wu et al., 26 May 2025). ASRR goes further toward explicit suppression by training under a “No-Thinking” prefix and then allowing a learned “Internal Self-Recovery Mechanism” to reintroduce reasoning only when needed (Zhang et al., 21 May 2025). This suggests that ARS is best treated as a family of selective-compute mechanisms rather than a single algorithmic pattern.

The same literature also separates efficiency-oriented suppression from safety-oriented suppression. Efficiency-oriented methods attempt to cut redundant reasoning while preserving correctness. Safety-oriented methods, such as reasoning-model unlearning, attempt to suppress reasoning traces that leak sensitive or harmful information even when the final answer has been erased (Wang et al., 15 Jun 2025). Both belong to ARS in the broad sense, but they optimize different failure modes.

2. Mechanisms and formal framing

The narrow ARS paper formulates reasoning generation for query $q$ as token generation

$$o = \{o_1, o_2, \ldots, o_T\}, \qquad o_t \sim \pi(\cdot \mid q, o_{<t}),$$

and states the objective as minimizing output length subject to preserving final-answer quality: $$\min_{\theta} \mathbb{E}[T] \quad \text{subject to} \quad \mathbb{E}[\mathcal{L}(f(o), y)] \leq \epsilon.$$ In that paper, suppression is selective rather than global: when the next token belongs to a trigger set such as “Wait,” “But,” or “Alternatively,” the decoder may block it and resample a non-trigger alternative if certainty is already high (Zheng, 29 Sep 2025).

A useful taxonomy suggested by the literature separates ARS mechanisms into pre-generation routing, training-time suppression, decoding-time suppression, and latent-space control. Pre-generation routing chooses a cheaper reasoning format before any long trace is emitted. Training-time suppression shapes the policy so that concise correct behavior is rewarded. Decoding-time suppression monitors reasoning online and stops, redirects, or attenuates it when continuation appears unproductive. Latent-space control intervenes directly on hidden states or neuron activations associated with bad reasoning states.

Method	Primary control point	Reported effect
ARS (Zheng, 29 Sep 2025)	Suppress reflection-trigger tokens at multi-checkpoint decoding time	Up to 53.0% token reduction, 46.1% latency reduction, and 57.9% energy reduction
ARM (Wu et al., 26 May 2025)	Pre-generation choice among Direct Answer, Short CoT, Code, and Long CoT	Average token savings of about 30%, and up to 70% on easy tasks
ASRR (Zhang et al., 21 May 2025)	No-Thinking prompting plus accuracy-aware length reward	Up to 32.5% reasoning-budget reduction on 1.5B and 25.7% on 7B with minimal pass@1 loss
ASAG (Li et al., 13 Jun 2026)	Attention-state early exit, logits injection, and jump prompting	On Qwen3-8B, average accuracy improves by 3.2% while tokens fall by nearly 40%
Ada-RS (Ge et al., 23 Feb 2026)	Training-signal filtering with adaptive length-penalized reward	Up to 80% lower average output tokens and up to 95% lower thinking rate

The mechanisms differ in where suppression is applied. ARM and ARM2 act before long reasoning begins; ASRR, Ada-RS, and ARLCP mainly alter the policy through RL; ARS and ASAG intervene during generation; AdaRAS and ActivationReasoning act on internal activations and can be read as suppressing error-associated or policy-violating latent states (Xie et al., 9 Oct 2025, Dong et al., 27 Jan 2026, Helff et al., 21 Oct 2025).

3. Training-time suppression and adaptive routing

A major line of ARS work replaces fixed long-CoT policies with learned adaptive policies. ARM trains a model on 10.8K AQuA-Rat questions paired with four reasoning formats, then applies Ada-GRPO to prevent format collapse toward Long CoT. Standard GRPO is reported to collapse toward Long CoT “early in training ($\sim$10 steps in our experiment),” so Ada-GRPO rescales rewards as

$r_i' = \alpha_i(t)\, r_i,$

with $\alpha_i(t)$ large for underrepresented formats early in training and decaying to $1$ by the end. The resulting model preserves performance while cutting pass@8 token usage by 33.6% at 3B, 32.5% at 7B, and 31.1% at 14B relative to the Long-CoT-dominant GRPO baseline; on easy tasks such as CSQA and OBQA, token reductions exceed 70% (Wu et al., 26 May 2025).

ARM2 extends the same idea to multimodal reasoning and executable code. Its GRPO-alp objective combines format encouragement with a relative length penalty,

$$r_i'' = \beta_i(t)\, r_i', \qquad \beta_i = \exp\left( - \lambda \cdot \frac{l_i - l_{\min}}{l_{\max} - l_{\min} + \epsilon} \right),$$

so longer responses are downweighted relative to shorter responses in the same group. ARM2 supports Direct Answer, Short CoT, Code-Text, Code-Exec, and Long CoT, and reports token usage reduced by over 70% on average while maintaining performance comparable to GRPO-trained long-reasoning baselines (Xie et al., 9 Oct 2025). In ARS terms, this is suppression by modality substitution as much as by shortening.

ASRR adopts a more explicit suppression-first stance. It trains under a No-Thinking prefix and uses an accuracy-aware dynamic length penalty. The reward is

$$\mathcal{R}_i = \mathbb{I}(y_i = \hat{y}_i) - \alpha \cdot \mathcal{O}_i,$$

where the overlong ratio $\mathcal{O}_i$ measures excess length relative to the shortest correct output in the group, and $$o = \{o_1, o_2, \ldots, o_T\}, \qquad o_t \sim \pi(\cdot \mid q, o_{<t}),$$0 becomes nonzero only after group accuracy exceeds a threshold $$o = \{o_1, o_2, \ldots, o_T\}, \qquad o_t \sim \pi(\cdot \mid q, o_{<t}),$$1 (Zhang et al., 21 May 2025). This prevents the common failure mode of optimizing for short but wrong traces. Relative to GRPO, ASRR reports 32.5% shorter generations on 1.5B and 25.7% shorter generations on 7B, with pass@1 drops of only 1.2 and 0.6 points.

Two later methods generalize training-time ARS beyond fixed math settings. Ada-RS treats selective thinking as a training-signal selection problem. For each context $$o = \{o_1, o_2, \ldots, o_T\}, \qquad o_t \sim \pi(\cdot \mid q, o_{<t}),$$2, it scores sampled completions with an adaptive length-penalized reward

$$o = \{o_1, o_2, \ldots, o_T\}, \qquad o_t \sim \pi(\cdot \mid q, o_{<t}),$$3

then uses stochastic rejection sampling to retain only high-reward candidates or preference pairs (Ge et al., 23 Feb 2026). ARLCP instead penalizes a specific subtype of overthinking—reflection markers such as “wait,” “hmm,” and “alternatively”—through a reflection penalty coordinated with a general length penalty. On DeepSeek-R1-Distill-Qwen-1.5B it reports a 53.1% average response-length reduction with a 5.8% accuracy gain, and on 7B a 35.0% reduction with a 2.7% gain (Yu et al., 12 Feb 2026).

4. Decoding-time stopping, suppression, and redirection

The namesake ARS method is purely inference-time and training-free. During generation it inserts checkpoints, probes a tentative answer from the current partial reasoning, computes a certainty score, tracks certainty history, and converts certainty plus a trend estimate into a suppression probability. If the next token is in the trigger set $$o = \{o_1, o_2, \ldots, o_T\}, \qquad o_t \sim \pi(\cdot \mid q, o_{<t}),$$4 and suppression fires, that token is rejected and a non-trigger alternative is sampled (Zheng, 29 Sep 2025). The method is evaluated on GSM8K and MATH500 with Qwen2.5-Math-1.5B-Instruct, Qwen2.5-Math-7B-Instruct, and DeepSeek-R1-Distill-Qwen-7B, and the best reported reductions reach 53.0% in tokens, 46.1% in latency, and 57.9% in energy.

ASAG replaces certainty-only stopping with a joint criterion based on confidence and attention entropy. At action transition points such as “Wait” or “Alternatively,” it probes an intermediate answer, computes confidence $$o = \{o_1, o_2, \ldots, o_T\}, \qquad o_t \sim \pi(\cdot \mid q, o_{<t}),$$5, aggregates deep-layer attention entropy $$o = \{o_1, o_2, \ldots, o_T\}, \qquad o_t \sim \pi(\cdot \mid q, o_{<t}),$$6, and defines an entropy-variation rate

$$o = \{o_1, o_2, \ldots, o_T\}, \qquad o_t \sim \pi(\cdot \mid q, o_{<t}),$$7

At the first ATP, it exits if $$o = \{o_1, o_2, \ldots, o_T\}, \qquad o_t \sim \pi(\cdot \mid q, o_{<t}),$$8; at later ATPs, it exits only if both $$o = \{o_1, o_2, \ldots, o_T\}, \qquad o_t \sim \pi(\cdot \mid q, o_{<t}),$$9 and $$\min_{\theta} \mathbb{E}[T] \quad \text{subject to} \quad \mathbb{E}[\mathcal{L}(f(o), y)] \leq \epsilon.$$0, with $$\min_{\theta} \mathbb{E}[T] \quad \text{subject to} \quad \mathbb{E}[\mathcal{L}(f(o), y)] \leq \epsilon.$$1 and $$\min_{\theta} \mathbb{E}[T] \quad \text{subject to} \quad \mathbb{E}[\mathcal{L}(f(o), y)] \leq \epsilon.$$2 in the reported setup (Li et al., 13 Jun 2026). If confidence is low but entropy indicates convergence, ASAG performs logits injection rather than hard stopping; if entropy indicates non-convergence and attention remains trapped on previous reasoning chunks, it injects a jump prompt. On Qwen3-8B the paper reports nearly 40% token reduction together with a 3.2% average accuracy gain across reasoning tasks.

STARS is adjacent rather than identical. It detects “Cognitive Pivots” from hidden-state L2 spikes,

$$\min_{\theta} \mathbb{E}[T] \quad \text{subject to} \quad \mathbb{E}[\mathcal{L}(f(o), y)] \leq \epsilon.$$3

uses cosine alignment and recurrence analysis to distinguish benign transitions from loop onset, and injects either a “Shifting Suffix” or a “Loop Breaker Suffix” (Lee et al., 30 Jan 2026). Its intervention is chiefly steering, not explicit stopping, but it is ARS-relevant because it shows that naive spike-based early exit causes large token reductions at substantial accuracy cost, whereas diagnosed redirection can improve both accuracy and efficiency.

5. Latent-space suppression, unlearning, and safety

One ARS branch targets harmful internal reasoning rather than redundant length. $$\min_{\theta} \mathbb{E}[T] \quad \text{subject to} \quad \mathbb{E}[\mathcal{L}(f(o), y)] \leq \epsilon.$$4 starts from the observation that conventional unlearning can erase final answers while leaving sensitive knowledge inside the chain of thought. Its key contribution is an “unthinking” loss that applies representation misdirection not only to forget-set prompts but also to reasoning traces elicited from segments of those prompts: $$\min_{\theta} \mathbb{E}[T] \quad \text{subject to} \quad \mathbb{E}[\mathcal{L}(f(o), y)] \leq \epsilon.$$5 This is combined with standard RMU and a CoT-preservation term on benign reasoning traces (Wang et al., 15 Jun 2025). The paper’s RT-UA metric counts reasoning traces in leakage classes C2–C4 and treats only C1 as successful unthinking, making trace suppression itself the evaluation target.

AdaRAS moves the control point deeper into MLP neurons. It scores neuron $$\min_{\theta} \mathbb{E}[T] \quad \text{subject to} \quad \mathbb{E}[\mathcal{L}(f(o), y)] \leq \epsilon.$$6 by the mean activation difference between correct and incorrect reasoning traces,

$$\min_{\theta} \mathbb{E}[T] \quad \text{subject to} \quad \mathbb{E}[\mathcal{L}(f(o), y)] \leq \epsilon.$$7

retains only neurons whose average activation changes sign between good and bad traces, selects the top $$\min_{\theta} \mathbb{E}[T] \quad \text{subject to} \quad \mathbb{E}[\mathcal{L}(f(o), y)] \leq \epsilon.$$8, and adds a sparse signed steering vector inside the MLP block only when a failure predictor says the current prompt is likely to fail (Dong et al., 27 Jan 2026). The paper reports gains of more than 13% on AIME-24 and AIME-25. In ARS terms, this is adaptive suppression of error-associated latent activation regimes rather than suppression of visible tokens.

ActivationReasoning is not itself an ARS method, but it provides a substrate for activation-level suppression. It represents latent concepts as sparse autoencoder features, activates propositions when concept scores exceed thresholds, reasons over those propositions with explicit rules, and can steer the hidden state with a signed decoder-direction update

$$\min_{\theta} \mathbb{E}[T] \quad \text{subject to} \quad \mathbb{E}[\mathcal{L}(f(o), y)] \leq \epsilon.$$9

Because the paper explicitly states that $\sim$0 can promote or suppress concept $\sim$1, it supplies the components needed for proposition-level ARS: concept thresholds, rule gating, and negative-$\sim$2 latent suppression (Helff et al., 21 Oct 2025).

6. Empirical regularities, limitations, and terminological boundaries

Across the literature, several regularities recur. First, easy tasks often suffer from fixed Long CoT: ARM reports that Direct Answer, Short CoT, or Code frequently suffice on commonsense and symbolic problems, while hard competition math triggers Long CoT more often (Wu et al., 26 May 2025). Second, complete no-thinking is too blunt. ASRR shows that No-Thinking reduces token usage sharply but loses substantial pass@1 on hard tasks unless the model can selectively recover reasoning (Zhang et al., 21 May 2025). Third, stronger suppression is not uniformly better: ARM2’s higher length penalties reduce tokens more aggressively but hurt AIME and GEO3K, and ARLCP likewise reports a standard trade-off in which overly large $\sim$3 harms accuracy (Xie et al., 9 Oct 2025, Yu et al., 12 Feb 2026).

The main limitations are equally consistent. Most methods reduce visible generated reasoning rather than hidden compute. ARM, ASRR, Ada-RS, and ARLCP all optimize output length, thinking rate, or reflection markers, but do not establish that internal computation has been equivalently reduced (Ge et al., 23 Feb 2026, Yu et al., 12 Feb 2026). Confidence-based online stopping can be brittle because models can be overconfident on hard tasks and underconfident on easy ones; this motivates ASAG’s use of attention entropy and STARS’s use of latent spikes rather than confidence alone (Li et al., 13 Jun 2026, Lee et al., 30 Jan 2026). Several methods are also domain-concentrated: the core evidence base remains strongest in math, coding, or synthetic tool-use environments (Zheng, 29 Sep 2025, Ge et al., 23 Feb 2026).

A further boundary concerns safety and domain specialization. TARS shows that adaptive reasoning allocation can improve the safety–refusal trade-off by spending more compute on ambiguous prompts, but it does not implement explicit suppression or no-reasoning gating at inference time (Kim et al., 1 Jul 2025). Emotion-o1 introduces adaptive reasoning depth control and suppression of repetitive logic for emotion, sentiment, humor, and sarcasm, but the paper does not report reasoning length, token usage, or latency metrics, and its reward mechanics are not fully specified (Song et al., 28 May 2025). These works enlarge the scope of ARS-like ideas while also showing that not every “adaptive reasoning” paper is an explicit suppression method.

Finally, the acronym itself is overloaded. In “Harnessing Reasoning Trajectories for Hallucination Detection via Answer-agreement Representation Shaping,” ARS denotes a detector-training method that perturbs the trace-boundary embedding, groups answer-preserving and answer-changing counterfactuals, and learns a contrastive representation for hallucination detection (Zhang et al., 24 Jan 2026). That work is detection-oriented and does not suppress reasoning at inference time. In current usage, therefore, “Adaptive Reasoning Suppression” is best reserved for methods that actively reduce, redirect, or unlearn unnecessary or harmful reasoning, whether by routing, reward shaping, decoding control, or latent-state intervention.