Sticker-TTS: Efficient History-Aware Reasoning

• Sticker-TTS is a test-time scaling framework that extracts, refines, and reuses compact reasoning artifacts ('stickers') to capture essential problem-solving conditions.
• It employs a three-model architecture—Sticker Extractor, Modifier, and Utilizer—to iteratively refine answers through structured historical information and majority voting.
• Evaluated on math benchmarks like AIME-24 and AIME-25, the framework demonstrates enhanced accuracy and efficiency by leveraging distilled intermediate reasoning.

Sticker-TTS is a test-time scaling framework for large reasoning models that is designed to utilize historical experience during iterative inference by extracting, refining, and reusing compact abstractions called “stickers” (Chen et al., 5 Sep 2025). In this usage, a sticker is not a visual asset but a distilled, structured abstract of the key conditions and strategy-driving steps in a reasoning process. The framework coordinates three collaborative LRMs—a Sticker Extractor, a Sticker Modifier, and a Sticker Utilizer—over multiple rounds, followed by majority voting on the answers. It was introduced to address the limitations of prevailing test-time scaling methods, especially redundant independent sampling and the direct reuse of full prior traces, and was evaluated on mathematical reasoning benchmarks including AIME-24, AIME-25, and OlymMATH (Chen et al., 5 Sep 2025).

1. Terminological scope and conceptual definition

In Sticker-TTS, the term sticker denotes a reasoning artifact rather than an image. It is defined as a distilled, structured abstract of the key conditions and strategy-driving steps that underpin a reasoning process for a given problem, and it contains a concise restatement of the question together with atomic, unambiguous, and indispensable conditions (Chen et al., 5 Sep 2025). This definition positions the sticker between two insufficient extremes: the full reasoning trace, which may be too verbose for efficient reuse, and the final answer alone, which omits the intermediate constraints and observations needed for correction or refinement.

The framework belongs to the literature on test-time scaling, where additional inference-time computation is used to improve performance. Its central claim is that historical attempts should not merely be accumulated or voted over; they should be compressed into high-signal abstractions that can scaffold later reasoning rounds. This suggests a shift from sampling-centric scaling toward history-aware iterative optimization.

A useful contextual distinction is that the name overlaps lexically with the sticker-generation literature, including text-to-sticker synthesis and animated sticker generation, but the 2025 Sticker-TTS framework is a reasoning system rather than a generative vision model (Chen et al., 5 Sep 2025). A plausible implication is that the name emphasizes compactness and reuse: the sticker functions as a concise, portable unit of prior experience.

2. Motivation and problem formulation

Sticker-TTS was motivated by two limitations in dominant inference-time reasoning strategies. First, multiple independent samplings and majority voting (self-consistency) do not leverage lessons from previous attempts and often repeat similar errors. Second, feeding entire previous solutions into the next query can produce either information overload or loss: verbose historical context may obscure key points, while single-step outputs may omit the intermediate cues necessary for improvement (Chen et al., 5 Sep 2025).

The framework addresses these issues by extracting only the essential reasoning anchors from historical attempts. In the formulation given for Sticker-TTS, those anchors are intended to be compact, structured, and task-relevant, allowing the model to preserve critical insight without burdening context length. The stated objective is history-aware, efficient, and focused test-time reasoning.

The evaluation domain is mathematical reasoning. The benchmarks explicitly reported are AIME-24, AIME-25, and OlymMATH-EN-EASY, and the metrics include Pass@1 and Cons@N, where the latter measures the effect of multi-round or multi-sample majority voting (Chen et al., 5 Sep 2025). The choice of these datasets indicates that the framework was intended for problems where intermediate constraints and strategic pivots matter, and where repeated independent attempts may plateau without a mechanism for reusing prior failures.

3. Stickers as distilled historical experience

The core abstraction in Sticker-TTS is the sticker itself. A sticker is extracted from a prior reasoning trace and is meant to capture the strategic essence of the attempt, including critical conditions, indispensable observations, and, where relevant, bottlenecks or errors (Chen et al., 5 Sep 2025). It is neither a chain-of-thought transcript nor an answer token; rather, it is a structured intermediate object optimized for transmission across reasoning rounds.

The paper organizes sticker use into three operations: extraction, refinement, and utilization. During extraction, a sticker is produced from the previous trace and the question. During refinement, the sticker is checked for flaws, corrected, and supplemented with missing conditions. During utilization, the revised sticker is supplied to the solution generator together with the original question and possibly the previous answer, so that the next trace is explicitly conditioned on prior distilled experience (Chen et al., 5 Sep 2025).

This design imposes several representational constraints. Stickers are described as atomic, essential, and unambiguous, and the prompting emphasizes minimalism and clarity. These properties are not incidental: they operationalize the claim that the useful residue of an unsuccessful or partial attempt is often much smaller than the trace that produced it. A plausible implication is that Sticker-TTS treats reasoning improvement as an information-selection problem as much as an inference problem.

4. Three-model architecture and recurrent inference loop

Sticker-TTS comprises three interrelated modules, each fine-tuned for a dedicated role: the Sticker Extractor $E$, the Sticker Modifier $M$, and the Sticker Utilizer $U$ (Chen et al., 5 Sep 2025). The extractor takes the previous reasoning trace $T^{(k-1)}$ and question $Q$ as input and outputs a structured sticker $s^{(k)}$. The modifier takes the sticker and the question and outputs a refined sticker $s^{(k)'}$. The utilizer then consumes the refined sticker, the original question, and the previous answer $A^{(k-1)}$ to produce a new reasoning trace $T^{(k)}$ and a new answer $A^{(k)}$.

The iterative procedure is specified as follows:

$M$2

This recurrent organization differs from self-consistency because later rounds are not independent draws. They are conditioned on structured summaries of prior attempts. It also differs from naïve iterative feedback because the transmitted object is a compressed abstraction rather than the full trace.

The division of labor is explicit. The extractor identifies what mattered, the modifier decides what must be repaired or added, and the utilizer transforms the revised abstraction back into a full reasoning attempt. This modularity is one of the framework’s defining characteristics. It supports the claim that test-time scaling can be decomposed into retrieval of salient historical information, correction of that information, and renewed task execution.

5. Two-stage optimization strategy

The framework uses a two-stage optimization strategy combining imitation learning with self-improvement (Chen et al., 5 Sep 2025). In Stage 1: Imitation Learning via Knowledge Distillation, the objective is to align the extractor, modifier, and utilizer with strong teacher behavior. The training pipeline described in the data uses math problems and “good” reasoning traces from teacher models such as DeepSeek-R1 and OpenAI o3-mini. The teachers are used to extract stickers, simulate correction scenarios, and generate improved traces, after which each module is fine-tuned on paired, filtered, and labeled data.

In Stage 2: Self-Improvement (Bootstrapping), the full Sticker-TTS loop is run on a large math problem set, identified as OpenThoughts. The resulting traces are filtered according to solution validity, and two kinds of examples are retained: error-to-correct cases, in which an incorrect original trace is successfully improved, and correct-to-correct cases, in which an already correct trace is further refined (Chen et al., 5 Sep 2025). The data selection procedure also favors examples requiring non-trivial correction, including cases with multiple failed prior iterations.

This training design is tightly coupled to the framework’s intended inference behavior. Rather than optimizing a single monolithic reasoner, Sticker-TTS trains specialized modules for abstraction, verification, and reapplication. A plausible implication is that the framework treats iterative reasoning as a learned workflow rather than a prompt-level heuristic.

6. Empirical results, ablations, and relation to sticker-generation research

Sticker-TTS was evaluated against self-consistency, reinforcement-learning-enhanced models, and other test-time scaling frameworks including Think-Twice and LeaP-T (Chen et al., 5 Sep 2025). The reported results state that it outperforms all tested strong baselines on the targeted mathematical reasoning benchmarks. For a 7B parameter model under Cons@20, the reported values are:

Benchmark	Sticker-TTS (Stage 2)	Comparison examples
AIME-24	83.33%	Skywork-OR1 76.67%, Light-R1 76.67%, LeaP-T 80.00%
AIME-25	63.33%	Skywork-OR1 63.33%, Light-R1 53.33%, LeaP-T 60.00%
OlymMATH	80.00%	Skywork-OR1 79.00%, Light-R1 65.00%, LeaP-T 62.00%

The paper further states that scaling to 32B models further increases gains, and that Sticker-TTS can reach or exceed the Cons@64 results of other baselines using only Cons@20-equivalent compute, which is attributed to reduced context and more focused reasoning per round (Chen et al., 5 Sep 2025). The efficiency claim is therefore not merely about accuracy; it concerns the use of inference budget under structured historical reuse.

The ablation studies support the role of each module. Removing the sticker extraction phase causes performance to drop, and removing error-correcting sticker refinement degrades performance further. A full ablation, removing stickers entirely, produces the most severe loss (Chen et al., 5 Sep 2025). The data also reports that deeper iteration improves performance, with Cons@N on OlymMATH rising from 56.67% at $M$0 to 83.33% at $M$1. These results reinforce the framework’s central thesis that distilled historical experience compounds over multiple rounds.

The name Sticker-TTS can be confused with the visual sticker literature. That literature includes “Text-to-Sticker: Style Tailoring Latent Diffusion Models for Human Expression” (Sinha et al., 2023), which studies text-to-sticker image generation via Style Tailoring for latent diffusion models, and “VSD2M: A Large-scale Vision-language Sticker Dataset for Multi-frame Animated Sticker Generation” (Yuan et al., 2024), which studies animated sticker generation and proposes the Spatial Temporal Interaction (STI) layer. These works use “sticker” in the conventional visual sense. By contrast, Sticker-TTS uses “sticker” as a structured reasoning abstraction. The overlap is terminological rather than methodological. A plausible implication is that the 2025 framework repurposes the notion of a sticker to denote a compact, expressive, and reusable unit of content, but in symbolic reasoning rather than image synthesis.