L2S Reasoning: Balancing Accuracy & Efficiency

• L2S is a framework that combines long-chain reasoning accuracy with concise output by merging System 1’s speed and System 2’s thoroughness.
• It employs techniques like training-free model merging, dynamic test-time trimming, and learned gating to optimize the trade-off between correctness and brevity.
• Empirical results show up to 55% reduction in token count with maintained or improved accuracy, highlighting practical efficiency gains in reasoning models.

Long-to-Short (L2S) denotes a class of methods that seek to preserve the problem-solving capability of long-chain-of-thought reasoning while reducing the number of generated tokens. In the large-language-model literature, the term is used for the correctness–brevity trade-off that arises when “System 2” reasoning models achieve strong performance through extended, deliberate reasoning but incur substantial inference-time cost (Iacobelli et al., 7 Apr 2026). The central objective is to maintain high accuracy with shorter reasoning traces, either by merging fast and slow checkpoints, dynamically trimming generation at test time, or training models to emit short or long reasoning paths conditionally (Wu et al., 26 Mar 2025, Han et al., 7 Sep 2025, Luo et al., 28 May 2025).

1. Definition and problem formulation

In L2S reasoning, two checkpoints of the same architecture are given: $\theta_{\mathrm{S1}}$ (“System 1,” fast but inaccurate, short CoT) and $\theta_{\mathrm{S2}}$ (“System 2,” slow but accurate, long CoT), both initialized from a common pre-trained $\theta_0$ (Iacobelli et al., 7 Apr 2026). The goal is to construct a merged or otherwise controlled model that maximizes accuracy while minimizing reasoning-trace length.

The bi-objective formulation given for Evo-L2S defines accuracy as Pass@1 over a benchmark $\mathcal D$ of $n$ problems,

$\Acc(\theta_M)=\frac1n\sum_{i=1}^n c_i, \quad c_i=\mathbf1\{\text{model answer to problem %%%%43%%%% is correct}\},$

and trace length as the mean number of generated tokens,

$\Len(\theta_M)=\frac1n\sum_{i=1}^n \ell_i, \quad \ell_i=\text{number of tokens generated on problem %%%%44%%%%}.$

The L2S target is therefore the bi-objective minimization

$\min_{\theta_M} F(\theta_M)=\bigl[-\Acc(\theta_M),\;\Len(\theta_M)\bigr].$

This makes explicit that correctness and brevity are conflicting objectives rather than a single scalar metric (Iacobelli et al., 7 Apr 2026).

A closely related formalization appears in EDIT, where a fixed pre-trained model $\mathcal M$ produces sampled reasoning chains for each problem, and the L2S objective is to find a sampling or generation strategy $\Gamma^*$ that maximizes accuracy subject to a hard or soft bound on average correct-chain length $\theta_{\mathrm{S2}}$0 (Han et al., 7 Sep 2025). That formulation emphasizes test-time control rather than parameter-space interpolation, but it encodes the same correctness-versus-conciseness trade-off.

The motivating phenomenon is overthinking: reasoning-capable models may generate redundant or circular steps that add little to answer quality while inflating cost and latency (Wu et al., 26 Mar 2025, Luo et al., 28 May 2025, Han et al., 7 Sep 2025). In this sense, L2S is not merely compression of output text; it is an attempt to recover shorter reasoning paths without discarding the robust behavior associated with long-form reasoning.

2. Conceptual framework: System 1, System 2, and Pareto trade-offs

The L2S literature consistently frames reasoning models through a System 1/System 2 dichotomy. System 1 models produce short, intuitive answers directly and are efficient but limited on complex tasks; System 2 models generate full chains of thought, including self-critique and iterative corrections, and are more accurate on hard problems but expensive at inference (Wu et al., 26 Mar 2025). L2S seeks a regime in which a model is concise like System 1 but accurate like System 2.

This framing leads naturally to Pareto-optimal reasoning. Evo-L2S treats L2S as a true bi-objective optimization problem and searches for a Pareto front of merged models rather than a single fixed merge coefficient (Iacobelli et al., 7 Apr 2026). EDIT likewise seeks a near-Pareto-optimal point by locating the smallest constraint $\theta_{\mathrm{S2}}$1 that still yields a stable, correct answer under constraint-guided generation (Han et al., 7 Sep 2025). AutoL2S approaches the same frontier through conditional generation: the model decides whether to take a short or long reasoning path based on the learned likelihood of a special <EASY> token versus <Long Trigger> (Luo et al., 28 May 2025).

The broader empirical study of model merging underscores that L2S is best understood as a balance between reasoning depth and practical efficiency. That study describes L2S as distilling System 2’s robust reasoning into a format that is as concise as possible—ideally close to System 1—while preserving System 2’s accuracy (Wu et al., 26 Mar 2025). This suggests that L2S is less a single algorithm than a design problem spanning model combination, decoding control, and data annotation.

3. Model-merging approaches

A major L2S line of work uses training-free model merging. The basic setting involves a base pre-trained model $\theta_{\mathrm{S2}}$2, a System 1 fine-tuned model $\theta_{\mathrm{S2}}$3, and a System 2 fine-tuned model $\theta_{\mathrm{S2}}$4, with the goal of producing a merged $\theta_{\mathrm{S2}}$5 that is concise like $\theta_{\mathrm{S2}}$6 but accurate like $\theta_{\mathrm{S2}}$7 (Wu et al., 26 Mar 2025).

Task-vector-based merging represents each fine-tuned model by a task vector $\theta_{\mathrm{S2}}$8 and forms

$\theta_{\mathrm{S2}}$9

Reported variants include Average Merging, Task Arithmetic (TA), TIES-Merging, and DARE (Wu et al., 26 Mar 2025). Evo-L2S specializes this idea to the two-checkpoint L2S setting using Task Arithmetic,

$\theta_0$0

and then optimizes $\theta_0$1 as a genotype in a multi-objective search (Iacobelli et al., 7 Apr 2026).

SVD-based methods such as LoRE-Merging and Twin-Merging attempt to exploit low-rank structure in task vectors to reduce interference during merging (Wu et al., 26 Mar 2025). Activation-informed methods, including AIM and Sens-Merging, use activations or sensitivities on a calibration set to assign layerwise or parameterwise weights (Wu et al., 26 Mar 2025).

A later development, FIM-Merging, argues that the linear-output assumption behind uniform Task Arithmetic is systematically violated in L2S settings and provides a theoretical justification for layer-adaptive merging (Xia, 23 Mar 2026). Proposition 1 bounds merging error by a term proportional to the per-layer Hessian norm, and the paper uses diagonal Fisher Information Matrix estimates as a computable proxy via Fisher–Hessian equivalence near local optima (Xia, 23 Mar 2026). The method computes diagonal FIM using only random token inputs and assigns per-layer coefficients $\theta_0$2, making the procedure data-free in the sense of requiring no domain-specific calibration data (Xia, 23 Mar 2026).

The significance of these merging approaches is that they target L2S without additional training. The empirical study explicitly contrasts merging with supervised fine-tuning, reinforcement learning, and prompt engineering, describing merging as parameter-only and adding zero extra training compute beyond a handful of SVD or activation passes (Wu et al., 26 Mar 2025). Evo-L2S retains the training-free character but replaces brittle scalarized search with explicit Pareto optimization (Iacobelli et al., 7 Apr 2026).

4. Evolutionary merging and layer-adaptive merging

Evo-L2S formulates L2S reasoning as a multi-objective optimization challenge and uses evolutionary model merging to approximate the Pareto-optimal set of merge coefficients (Iacobelli et al., 7 Apr 2026). The implementation uses NSGA-II as implemented in PyMoo, with a one-dimensional genotype $\theta_0$3, non-dominated sorting, crowding-based selection, binary tournament, Simulated Binary Crossover, and Polynomial Mutation (Iacobelli et al., 7 Apr 2026). Because PyMoo treats all objectives as minimization, Evo-L2S evaluates

$\theta_0$4

on a subset $\theta_0$5 of benchmark problems (Iacobelli et al., 7 Apr 2026).

A key technical ingredient is entropy-based subset sampling. Full-benchmark evaluation is described as too expensive, so Evo-L2S fixes $\theta_0$6 and selects the top-50 benchmark items by Bernoulli entropy

$\theta_0$7

where $\theta_0$8 is the empirical solve rate of item $\theta_0$9 across a calibration pool of $\mathcal D$0 merges (Iacobelli et al., 7 Apr 2026). Under a threshold model, the expected number of pairwise distinctions contributed by problem $\mathcal D$1 is proportional to $\mathcal D$2, which induces the same ranking as $\mathcal D$3; the paper further reports that with just 50 items entropy sampling achieves Spearman $\mathcal D$4 with the full benchmark ordering (Iacobelli et al., 7 Apr 2026).

The reported results position Evo-L2S as a more robust alternative to arithmetic merging with fixed hyperparameters. Across 1.5B, 7B, and 14B parameter scales on six mathematical reasoning benchmarks, Evo-L2S can reduce the length of generated reasoning traces by over 50% while preserving, or even improving, the problem-solving accuracy of the original reasoning models (Iacobelli et al., 7 Apr 2026). At 1.5B scale, the paper reports System 2 at Acc = 45.5% and Len $\mathcal D$5 tokens, ACM-TA at Acc = 46.4% and Len $\mathcal D$6 tokens, and Evo-L2S at Acc = 50.1% and Len $\mathcal D$7 tokens (Iacobelli et al., 7 Apr 2026). At 7B, the corresponding figures are 59.2% and $\mathcal D$8, 57.9% and $\mathcal D$9, and 59.6% and $n$0 (Iacobelli et al., 7 Apr 2026). At 14B, the paper reports that ACM-TA is brittle, with Acc = 47.7% and Len $n$1 tokens, whereas Evo-L2S reaches Acc = 58.9% and Len $n$2 (Iacobelli et al., 7 Apr 2026).

FIM-Merging addresses a related issue from a different angle. It argues that layer-adaptive merging is justified because merging error scales with layerwise curvature and task-vector magnitude, and it reports that FIM-TIES achieves state-of-the-art performance on five out of six evaluation benchmarks on the 7B L2S benchmark (Xia, 23 Mar 2026). On MATH500, FIM-TIES is reported at 90.2 versus 84.0 for ACM-TIES, and on the 1.5B benchmark FIM-TIES reaches an average accuracy of 47.3 while reducing average response length by 91.9% relative to the long-CoT model (Xia, 23 Mar 2026). A plausible implication is that L2S merging has evolved from uniform arithmetic interpolation toward curvature-sensitive and Pareto-aware procedures.

5. Test-time trimming and learned gating

Not all L2S methods operate in parameter space. EDIT is a test-time scaling method that addresses the L2S problem by repeatedly sampling under different length constraints and jointly tracking answer confidence and chain-length statistics (Han et al., 7 Sep 2025). For a constraint $n$3, it samples a candidate set $n$4, computes the most confident answer $n$5, and summarizes the lengths of chains producing $n$6 using

$n$7

The method then performs a binary search with patience to locate the smallest constraint that still yields a stable answer (Han et al., 7 Sep 2025). The paper states that as constraints tighten, the distribution of reasoning lengths shifts left, but below a critical $n$8 the predicted answer distribution suddenly collapses to incorrect solutions; EDIT is designed to exploit this sharp phase transition (Han et al., 7 Sep 2025).

AutoL2S represents a training-based alternative. It is described as a dynamic and model-agnostic framework that trains an autoregressive LLM on paired long and short CoT data plus a special <EASY> token (Luo et al., 28 May 2025). For each question, it obtains a long CoT $n$9 from a strong teacher model and a short CoT $\Acc(\theta_M)=\frac1n\sum_{i=1}^n c_i, \quad c_i=\mathbf1\{\text{model answer to problem %%%%43%%%% is correct}\},$0 from a math-capable model with rejection sampling; questions are marked EASY if the short CoT yields the correct answer (Luo et al., 28 May 2025). Standard cross-entropy training over the concatenated annotated sequences is sufficient for the model to learn an implicit gate between <EASY> and <Long Trigger> (Luo et al., 28 May 2025). At inference time, decoding begins by choosing between those two tokens, and that choice determines whether short or long reasoning is generated (Luo et al., 28 May 2025).

These approaches differ materially from merging methods. EDIT is entirely test-time and does not update parameters (Han et al., 7 Sep 2025). AutoL2S requires supervised fine-tuning on annotated long/short reasoning traces (Luo et al., 28 May 2025). Model-merging approaches, by contrast, are training-free and operate directly on checkpoints (Wu et al., 26 Mar 2025, Iacobelli et al., 7 Apr 2026). The literature therefore treats L2S as a shared objective that can be pursued through at least three distinct mechanisms: post hoc checkpoint interpolation, adaptive decoding, and explicit supervised routing.

6. Empirical patterns, benchmarks, and limitations

The empirical basis for L2S in reasoning is dominated by mathematical benchmarks. Evo-L2S evaluates Qwen2.5-Math and DeepSeek-R1-Distill models at 1.5B, 7B, and 14B scales on GSM8K, MATH500, Minerva-Math, OlympiadBench, College-Math, and AIME24 (Iacobelli et al., 7 Apr 2026). The earlier model-merging study uses Qwen2.5-Math and DeepSeek-R1 at 1.5B, 7B, 14B, and 32B, with the same six benchmarks and metrics including accuracy, average response length, and reflection ratio (Wu et al., 26 Mar 2025). EDIT evaluates GSM8K, MATH500, and AIMO across seven models ranging from vanilla instruction-tuned to RL-enhanced reasoning and distilled variants (Han et al., 7 Sep 2025). AutoL2S reports results on Math500, GPQA, GSM8K, and Olympiad with 3B and 7B base LLMs (Luo et al., 28 May 2025).

Several recurring empirical patterns appear across these papers. First, significant compression is often possible without large accuracy loss. The merging study reports up to 55% length reduction while preserving or slightly improving accuracy (Wu et al., 26 Mar 2025). Evo-L2S reports over 50% reduction in generated reasoning length while preserving, or even improving, accuracy (Iacobelli et al., 7 Apr 2026). AutoL2S reports reductions of up to 57% without compromising performance (Luo et al., 28 May 2025). EDIT reports 20–50% shorter outputs while matching or slightly improving the best baseline accuracy (Han et al., 7 Sep 2025).

Second, performance depends strongly on model scale and method class. The model-merging study reports that small models can reduce length on easy tasks but suffer accuracy loss on hard ones, whereas very large models may resist compression and sometimes even lengthen output (Wu et al., 26 Mar 2025). Evo-L2S similarly shows different Pareto behavior across 1.5B, 7B, and 14B scales (Iacobelli et al., 7 Apr 2026). FIM-Merging reports especially strong gains at 1.5B and 7B (Xia, 23 Mar 2026).

Third, the literature does not treat short reasoning as a uniform truncation problem. Merged models are reported to exhibit adaptive response length: they think more on hard problems and less on easy ones (Wu et al., 26 Mar 2025). AutoL2S formalizes that adaptivity through an <EASY> gate (Luo et al., 28 May 2025). EDIT formalizes it through dynamic constraint search (Han et al., 7 Sep 2025). This suggests that successful L2S systems compress selectively rather than simply shortening every output.

The limitations are correspondingly method-specific. Model merging is sensitive to hyperparameters and calibration data, and current methods may need finer-grained per-layer or per-module strategies (Wu et al., 26 Mar 2025). Evo-L2S addresses brittle fixed-hyperparameter arithmetic methods, but its fitness evaluation still depends on subset selection, making entropy sampling essential for recovering the high-accuracy, moderate-compression region (Iacobelli et al., 7 Apr 2026). EDIT must split a fixed sample budget across multiple constraint trials, and its length statistic was chosen empirically (Han et al., 7 Sep 2025). AutoL2S assumes that the short CoT is correct when marked <EASY>, and its implicit gate may be unstable when the probabilities of <EASY> and <LongTrigger> are close (Luo et al., 28 May 2025).

7. Scope of the term and adjacent uses

Within the supplied literature, “L2S” is not unique to long-to-short reasoning. In speech generation, SLD-L2S denotes “Hierarchical Subspace Latent Diffusion for High-Fidelity Lip to Speech Synthesis,” where L2S stands for lip-to-speech rather than long-to-short reasoning (Liang et al., 12 Feb 2026). In astrophysics, “The long-short GRB connection” uses an L2S label for an evolutionary channel from long gamma-ray bursts to short gamma-ray bursts within the binary-driven hypernova model (Rueda et al., 2024). These usages are terminologically distinct from the reasoning literature.

For reasoning models specifically, however, Long-to-Short has acquired a relatively coherent technical meaning: preserving the robust behavior of long-chain-of-thought systems while reducing response length, inference cost, and overthinking (Iacobelli et al., 7 Apr 2026, Wu et al., 26 Mar 2025, Han et al., 7 Sep 2025, Luo et al., 28 May 2025). The field’s current methodological spectrum spans training-free model merging, layer-adaptive curvature-aware merging, test-time constraint-guided trimming, and supervised dynamic routing. A plausible implication is that L2S has become a unifying objective for efficiency-oriented reasoning research rather than a single canonical algorithm.