Self-prompted Chain-of-Thought (SP-CoT)

• Self-prompted Chain-of-Thought (SP-CoT) is a framework where LLMs recursively generate their own reasoning traces and synthetic data for training and evaluation.
• It employs multi-stage pipelines like self-generation, compositional chaining, and adaptive demonstration sampling to minimize human annotation bottlenecks.
• Advanced filtering techniques, including self-consistency and reward inference, ensure high-quality reasoning traces and improved performance over traditional methods.

Self-prompted Chain-of-Thought (SP-CoT) is a class of semi- or fully-automated techniques leveraging LLMs to generate not only their own reasoning traces (chains of thought) but also task instructions, synthetic prompt datasets, or diverse reasoning trajectories for downstream training or inference. Unlike standard chain-of-thought (CoT) prompting, which directly elicits stepwise reasoning at inference, SP-CoT repurposes the model’s own recursive or planning capacities to produce training data, prompt variants, or evaluation traces—either for self-supervision or for curation of large, high-quality datasets. Recent methods under this umbrella include synthesis-driven dataset construction, multi-hop reasoning example generation, self-sampled compressed CoTs by activation manipulation, and curriculum-based efficient reasoning finetuning.

1. Core Principles and Definitions

SP-CoT methods distinguish themselves from canonical CoT prompting by using the model’s own outputs as intermediate supervision or as ingredients for prompt/data generation. In CoT-Self-Instruct, LLMs analyze a pool of seed instructions, produce structured CoT reasoning plans about instruction patterns, and then generate new prompts reflecting similar quality and complexity, with optional curation via answer consistency or learned reward metrics (Yu et al., 31 Jul 2025). Similarly, the S³-CoT framework operationalizes SP-CoT as “self-sampling via activation steering,” wherein the LLM autonomously generates variable-length reasoning traces for each task, which are then filtered or self-validated in the absence of teacher guidance (Du et al., 2 Feb 2026). In multi-hop question answering (QA), SP-CoT can refer to pipelines where LLMs synthesize question chains, explanations, and answer validation in fully automated loops, thus generating training datasets and in-context messages for downstream LLM inference (Wang et al., 2023).

Common to all SP-CoT approaches is the removal or reduction of human annotation bottlenecks by reusing the LLM’s own generative capacity for both supervision and evaluation. This often involves additional curation or filtering, which can be model-driven (e.g., reward-model scores, answer agreement, or self-consistency).

2. Automated Dataset Construction and Prompt Generation

SP-CoT frameworks for dataset construction use automated, multi-stage LLM pipelines. For open-domain multi-hop QA, the SP-CoT pipeline generates entity-centric passages, derives QA quadruplets with explicit double-check and rationale validation, and composes multi-hop chains via systematic templates (Wang et al., 2023). The pipeline can be summarized in three stages:

1. Self-generation of QA quadruplets: Given topics (e.g., “composers”), the LLM generates named entities, synthetic passages, candidate answers, questions, explanations, and performs answer validation by re-asking the question on its own passage.
2. Compositional chaining: Putative 2-hop chains are composed by graph-template linking, filtering near-duplicate chains, reformulating some into binary (yes/no) forms, and converting decomposed chains into natural multi-hop questions via additional LLM demonstrations.
3. Adaptive in-context demonstration sampling: Encodings (via Sentence-BERT) of all generated questions form a pool; at inference, test questions are clustered with these and nearest (diverse, relevant) in-context demonstrations are selected for prompting the LLM.

CoT-Self-Instruct instantiates SP-CoT for synthetic prompt generation in both reasoning and instruction-following domains. The method samples m seed prompts, constructs few-shot, CoT-based prompt templates, and has the LLM “reason and plan” before proposing a synthetic prompt. For verifiable tasks, answer-consistency filtering is applied by sampling multiple LLM completions and rejecting cases where the majority answer disagrees with the reference; for non-verifiable tasks, a “Reward-Inference-Preference” (RIP) score is used via a reward model over K completions (Yu et al., 31 Jul 2025). Surviving prompts are then used for reinforcement learning (e.g., GRPO or DPO fine-tuning).

3. Self-Sampled Succinct Reasoning and Dual-Cognitive Compression

SP-CoT also includes frameworks where the LLM actively modulates its own reasoning trace length or style. S³-CoT (Self-Sampled Succinct CoT) introduces a mechanism for self-generating variable-length CoTs without teacher supervision (Du et al., 2 Feb 2026). The process involves:

• Activation steering: Identification of a “variable-length direction” (VL-D) in hidden space by contrasting mean residual stream activations between “long-CoT” and “short-CoT” prompts. Linear interpolation along VL-D at certain layers systematically produces concise or verbose CoTs for the same input.
• Data filtering: Retention of only those self-sampled CoTs whose answers match gold labels or are self-consistent across variants (>99% accuracy when using self-consistency alone).
• Dual-cognitive curriculum-based fine-tuning: The LLM is taught both a “System 1” mode (concise reasoning under prompts like “answer concisely”) and a “System 2” mode (detailed step-by-step reasoning). Supervision is staged with a curriculum: initial finetuning focuses on traces with length ratio close to 1, gradually introducing shorter traces as the model capacity increases.

This dual-cognitive arrangement enables LLMs to efficiently trade off reasoning length against inference speed and accuracy, mimicking human-like fast and slow thinking regimes.

4. In-Context Self-Prompted Reasoning and Adaptive Demonstration Sampling

Inference-time SP-CoT involves assembling in-context demonstration sets using self-generated QA chains and CoT explanations. For each test question, relevant and diverse in-context examples are assembled by clustering candidate question embeddings and retrieving the most similar entry within each cluster. This maximizes coverage of different semantic regions and prevents redundancy (Wang et al., 2023). During inference, prompts take the following form: for each prior demonstration, the CoT reasoning and final answer are serialized, followed by the target question with partial CoT, allowing the model to continue stepwise reasoning.

Adaptive demonstration selection is formally defined as follows:

1. Clustering: All candidate questions $\{d_j\}$ are embedded to obtain $\{u_j\}$; K-means clustering partitions the pool into K clusters.
2. Cluster-wise retrieval: For each cluster $C_k$, pick $j^* = \arg\max_{j \in C_k} \cos(u_j, u^*)$ where $u^*$ is the embedding for the current test question.
3. Demonstration assembly: Demo$(Q) = \{d_{j_1^*}, ..., d_{j_K^*}\}$.

This enables both high relevance to the test question (via cosine similarity) and high coverage/diversity (via clustering).

5. Quality Control, Curation and Filtering Metrics

SP-CoT methods employ automatic quality control procedures to mitigate hallucinations and ensure prompt utility. In CoT-Self-Instruct, curation includes answer-consistency for verifiable tasks (reject if majority answer over K completions disagrees with the reference) and RIP scores (minimum reward score across K completions above a domain-specific threshold) for non-verifiable tasks (Yu et al., 31 Jul 2025). A summary of curation criteria follows:

Domain	Metric	Acceptance Criterion
Verifiable	Answer-Consistency	$A^* = t$
Verifiable	Self-Consistency	$SC(p) \geq 0.5$
Non-verifiable	RIP Score	$R_{min}(p) \geq \tau_{RIP}$

SP-CoT for dataset generation in QA includes a “double-check” validation (the LLM must answer its own question correctly in context) and decomposition/rationale sufficiency (Wang et al., 2023). For activation-steered self-sampling, S³-CoT uses either answer agreement with gold or self-consistency across all variants.

6. Empirical Performance and Comparative Analysis

On open-domain multi-hop QA tasks, SP-CoT demonstrates superior performance over both zero-shot, Auto-CoT, and Manual-CoT generation paradigms (Wang et al., 2023). Table 1 (reconstructed for clarity):

Method	MSQ EM/F1	HotpotQA EM/F1	2Wiki EM/F1	CWebQ EM/F1	Avg EM/F1
Zero-Shot	3.1/7.3	22.4/30.0	18.7/21.7	31.6/37.5	19.0/24.1
Zero-Shot-CoT	5.0/8.8	22.6/29.6	24.3/27.1	30.3/36.2	20.6/25.4
Auto-CoT	8.1/13.6	26.1/36.3	26.2/30.2	29.9/38.4	22.6/29.6
Manual-CoT	12.3/19.2	32.4/43.7	27.7/34.6	36.6/43.0	27.3/35.1
SP-CoT	14.5/22.6	33.2/42.9	30.1/34.7	37.5/43.6	28.8/36.0

SP-CoT improves both Exact Match (EM) and F1, and produces reasoning traces with higher clarity, conciseness, directness, and intermediate-answer recall (49%) relative to preceding methods (32% for Auto-CoT, 15% for Zero-Shot-CoT).

In the domain of synthetic prompt generation for reasoning and instruction-following, CoT-Self-Instruct with answer-consistency filtering substantially outperforms both standard self-instruct and human-written prompts in downstream RL fine-tuning, as measured by accuracy on math benchmarks and winrate on instruction evaluations (Yu et al., 31 Jul 2025). S³-CoT reduces average CoT length by 17–40% with minimal loss (and often some gain) in accuracy across math and medical domains, demonstrating strong accuracy–efficiency trade-offs (Du et al., 2 Feb 2026).

7. Limitations and Extensions

SP-CoT methods inherit seed-pool biases when generating synthetic instructions or prompts, and their data quality depends on filtering thresholds or reward-model coverage (Yu et al., 31 Jul 2025). Certain ablations indicate that longer CoT plans yield better prompts; self-consistency and answer-consistency filtering outperform simple generation or short-CoT variants. Failure modes include coverage gaps when using narrow seed pools, limitations for very deep multi-hop chains, and reduced efficacy with non-instruction-finetuned LLMs (Wang et al., 2023).

Potential extensions include dynamic expansion of seed pools via external retrieval, adaptive thresholding of curation metrics, and recursive SP-CoT loops where newly trained models regenerate and refine seed prompts, subject to external or retrieval-based quality guarantees. S³-CoT’s “self-evolution” regime—relying solely on self-consistency for data filtering—achieves >99% precision and maintains high accuracy without gold labels, suggesting robust potential for fully autonomous SP-CoT variants (Du et al., 2 Feb 2026).

---

In summary, Self-prompted Chain-of-Thought encompasses a spectrum of frameworks that transform LLMs from passive responders into active agents capable of recursively improving their own reasoning, generating training data, and compressing their outputs for greater efficiency—all while preserving or enhancing quality relative to both manual and prior automated approaches (Wang et al., 2023, Yu et al., 31 Jul 2025, Du et al., 2 Feb 2026).