Long-Short CoT Fusion: Adaptive Reasoning

Updated 8 July 2026

Long-Short Chain-of-Thought Fusion is a family of techniques that integrate detailed long-form reasoning with concise short traces to balance depth and efficiency.
Data-centric methods like LS-Mixture SFT and R1-Compress demonstrate performance gains (e.g., 2.3% accuracy improvement) while significantly reducing token count.
Hierarchical and adaptive fusion approaches enable models to dynamically switch between long and short reasoning modes, minimizing redundancy and enhancing problem-solving.

Long-Short Chain-of-Thought Fusion denotes a family of methods that combine long and short reasoning regimes in order to retain the problem-solving benefits of extended chain-of-thought (CoT) while reducing verbosity, token cost, or overthinking. Across recent work, the term covers several distinct mechanisms: mixture-style supervised fine-tuning on paired long and short traces, selective compression of long traces into shorter but structured rationales, hierarchical systems that interleave short planning with long execution, latent internalization of long reasoning for short or thought-free inference, and adaptive policies that switch between short and long modes by task difficulty. The common motivation is consistent: long CoT can improve difficult reasoning, but unconstrained long traces often waste computation, induce redundancy, or degrade behavior on easier tasks (Yu et al., 6 May 2025).

1. Conceptual scope and defining dimensions

A useful starting point is the distinction between Short CoT and Long CoT developed in the long-CoT survey literature. Short CoT is characterized as a “shallow, linear reasoning process” with a limited number of logical nodes, whereas Long CoT relaxes those constraints through a larger reasoning budget, branching exploration, and revisitation or reflection. The survey formalizes this contrast as a shift from bounded linear chains to structures with deeper reasoning, extensive exploration, and feasible reflection, and it explicitly frames adaptive reasoning techniques as a future direction (Chen et al., 12 Mar 2025).

Within that conceptual frame, “fusion” is not a single architecture. In some papers it means explicit coexistence of short and long traces in training data; in others it means alternation between short planning and long execution; elsewhere it means using long reasoning only during training while deploying short or thought-free outputs at inference. Hunyuan-TurboS makes this broad interpretation explicit by describing an “adaptive long-short CoT fusion method” that trains a single model to “autonomously decide whether to use long or short CoT and determine the depth of reasoning based on the complexity of the problem” (Team et al., 21 May 2025).

This literature therefore treats long-short fusion less as a fixed model family than as a control problem over reasoning depth, branching, and representation. A plausible implication is that the central question is not whether long CoT or short CoT is universally better, but which components of long reasoning must remain explicit, which can be compressed, and when deeper deliberation should be invoked at all.

2. Data-centric mixture and compression approaches

One major line of work performs fusion at the data level. “Long-Short Chain-of-Thought Mixture Supervised Fine-Tuning” describes a training recipe that combines long CoT reasoning data with short counterparts obtained through “structure-preserved rewriting.” In the abstract-level characterization available for that paper, LS-Mixture SFT improves average accuracy by 2.3% across various benchmarks while reducing response length by approximately 47.61% relative to direct SFT, with the stated goal of transferring reasoning ability without inheriting teacher overthinking (Yu et al., 6 May 2025).

A more structural variant appears in DLCoT, which treats long CoT as a decomposable object rather than a flat sequence. DLCoT parses traces into macro-components, segments approach exploration and verification, clusters similar branches, removes redundancy while retaining at least one representative per cluster, and then reconstructs coherence. Its results show that structured shortening can outperform raw long-CoT distillation. For example, on R1-distilled Qwen2.5-14B, the DLCoT-multiall variant improves AIME2024 from 46.7 to 53.3 while reducing the reported token count from 28526.5 to 18825.0. Equally important, DLCoT-multiall-incorrectness performs worse, indicating that removing all erroneous intermediate states is harmful rather than beneficial (Luo et al., 20 Mar 2025).

R1-Compress occupies a related but distinct position. Instead of pruning at the whole-instance level or skipping tokens directly, it compresses long CoT at the chunk level: segment into chunks, rewrite each chunk into multiple shorter candidates, then perform inter-chunk search for a coherent sequence. On MATH500 with Qwen2.5-32B-Instruct, it reaches 92.4 accuracy with only a 0.6% drop relative to the Long-CoT baseline while reducing token usage by about 20%. The paper also reports that R1-Compress preserves about 78% of the reflections in Long-CoT on MATH500, which is relevant because reflection is treated there as part of long-CoT’s functional value rather than mere verbosity (Wang et al., 22 May 2025).

Taken together, these works define one core interpretation of long-short fusion: preserve the reasoning trunk and selected reflective structure, compress repetitions and redundant branches, and avoid the assumption that every visible error state should be removed.

3. Hierarchical and cooperative fusion mechanisms

A second line of work performs fusion during inference or joint generation, often by assigning different roles to short and long reasoning components.

Long $\otimes$ Short is the clearest thought-level example. It first chunks long CoT into thoughts, scores each thought by a metric combining effectiveness and efficiency, preserves important thoughts in long form, compresses the rest, and lets two specialized LLMs collaborate: a long-thought LLM for important thoughts and a short-thought LLM for the remainder. On Qwen2.5-7B, the reported Long-r4⊗Short-r2 model achieves MATH500 89.80, AIME24 46.67, AIME25 33.33, GPQA Diamond 47.97, AMC23 90.00, with average length 2,113, compared with the DeepSeek-R1-Distill-Qwen-7B baseline length of 24,566; the corresponding length reduction is about 91.4% (Ning et al., 17 May 2025).

MPPA addresses long reasoning from a planning-execution perspective. It segments long CoT into planning and execution steps, treats most reasoning errors as planning errors, samples multiple short plan rollouts at planning steps, aggregates them into a refined planning step, and then resumes full execution. The aggregation policy is implemented as a lightweight LoRA module, and the inference schedule triggers plan aggregation at token gaps of 256, 512, and 1024 depending on current position. This is an explicit local fusion of short planning with long execution rather than a global shortening method (Xiong et al., 13 Oct 2025).

UCoT shifts the fusion locus into latent space. A small compressor generates fixed-length “upfront thought” embeddings from a question, and a large executor consumes those embeddings to produce a shorter explicit CoT. The paper formalizes ordinary CoT compression with a target ratio $\alpha$ , then uses hidden-state placeholder positions to encode a dense reasoning representation $H_n$ . On GSM8K with Qwen2.5-7B-Instruct, UCoT reduces token usage by 50% while delivering performance 3.08% higher than the stated SOTA method (Li et al., 9 Oct 2025).

Method	Fusion locus	Reported effect
Long $\otimes$ Short	Long thoughts + short thoughts across two LLMs	About 91.4% length reduction on Qwen2.5-7B
MPPA	Short plan aggregation inside long execution	Improves long-CoT robustness via planning refinement
UCoT	Latent upfront embeddings + short executor CoT	50% token reduction on GSM8K

These mechanisms share a common design principle: long reasoning is treated as selectively valuable, not uniformly necessary. The fused system therefore preserves depth where it is high-yield—important thoughts, planning steps, or latent structural guidance—while shifting routine continuation into shorter forms.

4. Internalization, minimal grounding, and adaptive depth

A third cluster of papers moves fusion away from explicit long/short text pairing and toward internalization or adaptive mode selection.

3TF formulates the problem as “Thought-Training and Thought-Free inference.” Training uses explicit CoT-annotated data with the factorization

$P_\theta(t, a \mid x) = P_\theta(t \mid x)\, P_\theta(a \mid t, x),$

while inference suppresses the external reasoning field and uses

$P_\theta(a \mid x, t=\varnothing).$

Its central claim is that long explicit thought is useful as a training signal even when deployment uses short or answer-only outputs. On Qwen3-8B, for example, 3TF improves AIME24 from 50.0 in NoThinking mode to 76.6 while using 60.4% of Think-mode tokens (Wu et al., 5 Nov 2025).

Draft-Thinking develops a related but distinct idea: train a model to produce concise “draft” traces that retain only key inferences, then scale capability through progressive curriculum learning rather than post hoc shortening. Its MATH500 headline result is an 82.6% reduction in reasoning budget at a cost of only a 2.6% performance drop. The paper also introduces an adaptive prompt asking the model to choose between detailed reasoning and minimal draft reasoning based on question difficulty, which effectively turns reasoning depth into a model-selectable behavior (Cao et al., 28 Feb 2026).

The vision-centric study on concise grounding pushes the asymmetry further. Comparing Language CoT, Grounding CoT, Visual CoT, and the minimal grounded variant G-CoT-least, it finds that visual and longer CoT mainly accelerate convergence but do not raise the final performance ceiling, while concise grounded supervision generalizes better. In maze experiments, G-CoT-least starts from a higher performance level, converges faster than explicit grounding CoT, and generalizes more robustly across maze sizes (Du et al., 27 Nov 2025).

Hunyuan-TurboS provides the most explicit adaptive deployment claim. It reports average output tokens of 1207.8 on a 6k-prompt mixed evaluation set, compared with 2283.5 for DeepSeek-R1 and 2979.2 for Qwen3-235B-A22B, while stating that the model automatically activates a “no thinking” mode for simple questions and a “thinking” mode for complex problems (Team et al., 21 May 2025).

This group of results suggests that long-short fusion can also be realized without explicit side-by-side trace generation: by internalizing long reasoning during training, by learning concise draft-style reasoning as a native mode, or by training a single policy to switch between short and long behaviors online.

5. Theory, structural analysis, and limits of “more thinking”

Several papers analyze the long-short problem at the level of theory or reasoning structure rather than implementation.

“How does Chain of Thought decompose complex tasks?” models CoT as a decomposition of a large classification problem into smaller ones. Its central result is that there is an optimal local branching factor

$m^* = e^{d/2},$

and an optimal finite reasoning depth

$n^* = \frac{2}{d}\ln N.$

Below the threshold, “thinking more” is detrimental; above it, more depth helps only until effective local difficulty approaches $m^*$ , after which further depth worsens error (Nadgir et al., 10 Apr 2026).

“Let Me Think!” sharpens the sequential-versus-parallel contrast. On bridge-graph constructions, it shows that one sufficiently long CoT can be worth exponentially many short ones, because parallel short traces cannot compensate for missing sequential depth. Its empirical conclusion is equally important: parallel scaling becomes useful only after sequential scaling has already crossed a minimum threshold for nontrivial performance (Mirtaheri et al., 27 May 2025).

LCoT2Tree adds a diagnostic layer. By converting sequential long CoT into hierarchical reasoning trees, it shows that structure predicts correctness better than raw length. Across five model families, tree-based classifiers improve over length-only baselines by 3.89% to 8.27% on average, and the explainability analysis highlights four recurring failure motifs: over branching, step redundancy, direct reasoning, and skipped thinking. This directly undermines any account of fusion based solely on token counts (Jiang et al., 28 May 2025).

The survey literature is consistent with these theoretical and diagnostic results: long CoT is valuable because of deep reasoning, extensive exploration, and feasible reflection, but overthinking and reasoning-boundary effects imply that “more” is not a monotone resource (Chen et al., 12 Mar 2025).

A plausible synthesis is that successful long-short fusion requires bounded and structured escalation. Long reasoning helps when it resolves genuine local ambiguity, supports correction, or explores alternatives with semantic value; it hurts when it merely adds repeated validation, diffuse branching, or post-solution continuation.

Not every relevant paper is an explicit fusion method, but several clarify the boundary conditions of the area.

CAC-CoT is a compact-CoT synthesis framework rather than an explicit fusion architecture. It constrains synthetic traces with a fixed inventory of “correct” and “incorrect” connectors and uses standard SFT on the resulting data. On S1-Bench it achieves English Acc@5 $= 86.07$ with ART $\alpha$ 0, compared with much longer ART values for long-CoT baselines, while its training traces average 1843.43 tokens versus 9291.62 for s1K-1.1. The paper’s stated goal is to bridge System-2 and System-1 behavior by compactness rather than routing (Choi et al., 26 Aug 2025).

Shift-FFN is not a fusion method at all, but it addresses a failure mode that matters for the long side of any fusion pipeline: cyclical reasoning during long-CoT adaptation. By editing current token representations with the previous token before the FFN, it increases adjacent-token representation differences and lowers the rate of “Length Exceeded Percentage.” On Qwen2.5-7B with LoRA rank 256, LoRA+Shift-FFN reaches average accuracy 51.2 with $\alpha$ 1, compared with 50.4 and $\alpha$ 2 for standard LoRA at the same rank (Xu et al., 22 May 2025).

RCP-Merging addresses a different fusion problem—merging reasoning models with domain-specific models—but its asymmetry is instructive. It treats the reasoning model as prior and selectively admits incoming updates according to a reasoning preservation indicator derived from diagonal Fisher information. The resulting merge has the form

$\alpha$ 3

which strongly suggests that long-reasoning capability may need to be preserved as a protected parameter subspace rather than symmetrically averaged with other behaviors (Yang et al., 5 Aug 2025).

Framework of Thoughts contributes the infrastructure perspective. It formalizes reasoning as a dynamic execution graph and a reasoning graph, supports graph-modifying operations, parallel execution, intelligent caching, and optimization over prompts and hyperparameters. FoT is not itself a long-short fusion policy, but it supplies the exact control abstractions needed to implement one: start with a chain, branch only when needed, run verifiers concurrently, and optimize the short/long tradeoff under runtime and cost constraints (Fricke et al., 18 Feb 2026).

Across these papers, several unresolved questions recur. The literature still lacks a canonical difficulty estimator for deciding when short reasoning should escalate into long reasoning; it does not yet provide a universally accepted faithfulness criterion for compressed or latent traces; and it rarely offers calibrated compute-aware objectives that jointly optimize accuracy, latency, and explicit reasoning length. The evidence so far nonetheless converges on one point: long-short CoT fusion is best understood as selective preservation and adaptive deployment of reasoning structure, not as a simple preference for shorter or longer outputs.