Generative Knowledge Prompting

• Generative knowledge prompting is a prompt-based method that first elicits intermediate knowledge from a language model to enhance downstream reasoning tasks.
• It employs multi-stage processes—knowledge generation followed by answer synthesis and selection—to improve unsupervised commonsense question answering.
• Empirical results show notable gains (e.g., +16.8% on CommonsenseQA) and versatile applications in prompt synthesis, slot filling, affective computing, and 3D generation.

Searching arXiv for recent and foundational papers on generative knowledge prompting and closely related formulations. Generative knowledge prompting is a prompt-based methodology in which a generative model is used not only to produce a final answer or artifact, but first to surface intermediate knowledge, prompting strategies, or structured semantic scaffolds that are then reused downstream. In the literature, the term most directly denotes the two-stage commonsense reasoning procedure introduced in “Generated Knowledge Prompting for Commonsense Reasoning” (Liu et al., 2021), where a LLM first generates natural-language knowledge and then conditions answer prediction on that knowledge. Subsequent work generalizes the pattern across unsupervised commonsense question answering, prompt synthesis, knowledge graph construction, slot filling, affective computing, and 3D generation, but the shared mechanism remains the same: prompting is used to externalize latent knowledge from a pretrained model, and the generated intermediate content becomes an explicit conditioning signal for a later inference stage (Liu et al., 2021, Sun et al., 2022).

1. Conceptual definition and scope

Generative knowledge prompting differs from direct prompting in that the model is not asked only for the terminal output. Instead, the model is first prompted to generate some intermediate content that functions as knowledge, guidance, or a reasoning scaffold. In “Generated Knowledge Prompting for Commonsense Reasoning” (Liu et al., 2021), this is formulated as generating knowledge statements $K_q$ conditioned on a question $q$, and then predicting the answer using the augmented input $[k_m \Vert q]$. The paper explicitly characterizes the method as generating knowledge from a LLM and then providing that knowledge as additional input when answering a question (Liu et al., 2021).

Within commonsense reasoning, “TSGP: Two-Stage Generative Prompting for Unsupervised Commonsense Question Answering” (Sun et al., 2022) sharpens this formulation by separating three components: knowledge generation, answer generation, and answer prediction. The first stage produces free-form commonsense statements $K_q = \{k_1, k_2, \dots, k_m\}$, described as “a series of coherent short sentences” that fill knowledge gaps and clarify implicit intermediate reasoning steps. The second stage then generates pseudo-answers $S_q = \{s_1, s_2, \dots, s_n\}$ conditioned on both the question and the selected knowledge statement. The third stage selects the final answer by semantic comparison between pseudo-answers and answer choices (Sun et al., 2022).

This suggests a broad definition: generative knowledge prompting is a family of methods in which prompting is used to make implicit model knowledge explicit enough to be reused by a later stage. In the commonsense setting, the generated artifact is natural-language knowledge. In other settings, the generated artifact may instead be a prompt, a schema, a graph component, a rationale, or a latent conditioning signal. The common structure is still “generate knowledge first, then solve the task with that generated knowledge” (Liu et al., 2021, Sun et al., 2022).

2. Foundational formulation in commonsense reasoning

The foundational formulation appears in multiple-choice commonsense reasoning. In TSGP, the task is formalized as predicting

$\hat{a} = \arg\max_{a \in A} P(a \mid q). \tag{1}$

The setting is unsupervised: the model does not use labeled QA pairs for training. The paper argues that earlier unsupervised methods suffer from two main limitations: retrieval from external knowledge bases is narrow and dependent on indexing and matching, while fixed-type knowledge generation can elicit only predefined forms of knowledge and therefore generalizes poorly across domains and tasks (Sun et al., 2022).

Generated knowledge prompting was introduced earlier as a two-stage alternative to both direct prompting and external retrieval. In that formulation, the generator samples knowledge statements

$$K_q = \{k_m : k_m \sim p_G(k \mid q), \; m=1,\dots,M\},$$

and the inference model predicts

$\hat a = \arg\max_{a \in A_q} p_I(a \mid q, K_q).$

The actual integration rule tries multiple generated knowledge statements and chooses the answer supported by the strongest one: $$\hat a = \arg\max_{a \in A_q} \max_{0 \le m \le M} p_I(a \mid q_m),$$ where $q_m = [k_m \Vert q]$ (Liu et al., 2021).

Both papers frame the same central claim from slightly different angles. The 2021 work argues that large pretrained LLMs can serve as flexible sources of external knowledge without requiring a structured knowledge base (Liu et al., 2021). The 2022 work extends that view by insisting that knowledge elicitation and answer candidate generation should be separated, rather than collapsed into a single scoring step or a single answer-generation prompt (Sun et al., 2022).

A common misconception is that generative knowledge prompting is just chain-of-thought under another name. The papers do not define it that way. In the foundational commonsense papers, the intermediate text is not merely a reasoning trace but an explicit knowledge statement intended to support later inference. The model is prompted to produce facts, paraphrases, causal relations, social norms, or physical regularities, and only afterward is that generated material used for answer selection (Liu et al., 2021, Sun et al., 2022).

3. Canonical two-stage and three-stage architectures

The canonical architecture has at least two stages. In the original generated knowledge prompting framework, stage one generates knowledge and stage two integrates it into answer prediction (Liu et al., 2021). TSGP makes the pipeline three-part: generate knowledge, generate pseudo-answers from the knowledge, and then select the answer through semantic similarity (Sun et al., 2022).

In TSGP, knowledge generation uses prompts containing an instruction, demonstrations, and a placeholder for the new question. For SocialIQA, the instruction is “Generate some knowledge about the events in the input.” For CommonsenseQA and OpenBookQA, the instruction is “Generate some knowledge about the concepts in the input.” The paper states that the demonstrations are hand-written and task-style, and for implementation they are randomly sampled from the training set (Sun et al., 2022).

Given a question $q$0, the model samples $q$1 knowledge statements per question using nucleus sampling with $q$2, discarding duplicates and empty strings. Because the generated knowledge can be noisy, TSGP ranks candidate knowledge statements with point-wise mutual information: $q$3 and then selects

$q$4

The selected knowledge statement is then injected into an answer-generation prompt, which also contains an instruction, demonstrations, a knowledge placeholder, and a question placeholder (Sun et al., 2022).

The answer-generation stage is deliberately unconstrained by the provided answer choices. TSGP generates a set of pseudo-answers

$q$5

using GPT-2, nucleus sampling with $q$6, and sample size $q$7. The model is thus allowed to generate semantically plausible answer expressions that are not literal copies of answer choices, reducing dependence on surface-form overlap (Sun et al., 2022).

The answer-prediction stage then computes semantic support for each answer choice: $q$8 using SRoBERTa-large fine-tuned on NLI and $q$9. This shifts the final decision away from raw language-model likelihoods, which the paper argues are sensitive to word frequency, sentence structure, and length (Sun et al., 2022).

A plausible implication is that generative knowledge prompting became progressively more modular after the 2021 formulation. The earlier work shows that generated knowledge can be appended directly to the question (Liu et al., 2021). TSGP suggests that further decomposition—knowledge elicitation followed by pseudo-answer elicitation—can make latent reasoning more explicit and more robust under unsupervised conditions (Sun et al., 2022).

4. Prompt design, elicited knowledge types, and scoring mechanisms

Prompt design is central because the method relies on eliciting useful intermediate content without task-specific finetuning for knowledge integration. In the original generated knowledge prompting paper, each task uses five demonstrations per prompt. Example instructions include “Generate some numerical facts about objects” for NumerSense and “Generate some knowledge about the concepts in the input” for CSQA and QASC (Liu et al., 2021). The generated output is a natural-language knowledge statement rather than a direct answer.

TSGP adopts a similarly simple prompt structure but emphasizes that the generated knowledge is unrestricted in type. The paper contrasts this with earlier prompt-based methods that elicit only predefined forms such as information-seeking questions or relation-specific inferences. Its knowledge prompts are intended to generate unlimited types of commonsense knowledge, including paraphrase-like statements, causal relations, social norms, common physical facts, and analogy-like inferences (Sun et al., 2022).

The following table summarizes the main prompt-mediated components reported across the foundational commonsense formulations.

Component	Generated artifact	Reported mechanism
Generated knowledge prompting (Liu et al., 2021)	Natural-language knowledge statement	Generate $[k_m \Vert q]$0, append $[k_m \Vert q]$1, score answers
TSGP stage 1 (Sun et al., 2022)	Commonsense knowledge statement	Sample $[k_m \Vert q]$2, rank with $[k_m \Vert q]$3, keep best
TSGP stage 2 (Sun et al., 2022)	Pseudo-answer set $[k_m \Vert q]$4	Sample $[k_m \Vert q]$5, compare semantically to answer choices

The scoring mechanisms reflect two different views of how generated knowledge should be consumed. The 2021 paper uses the answer support induced by individual knowledge statements and finds that a max-based selection over generated knowledge works better than mixture-of-experts or product-of-experts aggregation (Liu et al., 2021). TSGP, by contrast, uses PMI to filter knowledge before answer generation, and then uses semantic similarity over many pseudo-answers to select among answer choices (Sun et al., 2022).

This suggests that generative knowledge prompting is not tied to a single downstream scoring rule. What defines the method is the explicit production of intermediate knowledge-like content, not whether that content is consumed by likelihood maximization, semantic voting, constrained decoding, or another scoring function.

5. Empirical performance and analytical findings

The foundational commonsense papers report that generated intermediate knowledge improves performance across several benchmarks. “Generated Knowledge Prompting for Commonsense Reasoning” (Liu et al., 2021) improves zero-shot and finetuned models on NumerSense, CommonsenseQA, CommonsenseQA 2.0, and QASC. Reported examples include 64.05 → 72.47 for zero-shot T5 on NumerSense, 39.89 → 47.26 for zero-shot T5 on CommonsenseQA, 70.2 → 73.03 for finetuned Unicorn on CommonsenseQA 2.0, and 76.74 → 80.33 for finetuned UnifiedQA on QASC (Liu et al., 2021).

TSGP evaluates on CommonsenseQA, OpenBookQA, and SocialIQA and reports the best unsupervised performance among compared methods on all three datasets. Reported improvements over the baseline are +16.8% on CommonsenseQA, +21.6% on OpenBookQA, and +8.7% on SocialIQA. It also improves over SEQA by +5.2%, +10.2%, and +4.0% on the same datasets, respectively (Sun et al., 2022).

The TSGP ablation isolates the contribution of each stage. On CommonsenseQA and OpenBookQA, the reported scores are:

Setting	CommonsenseQA	OpenBookQA
GPT2-XL baseline	32.3	22.8
+ Knowledge Generation	45.5	30.0
+ Answer Generation	43.9	42.0
+ Both (TSGP)	49.1	44.4

These results support the paper’s claim that both components matter and that the combined framework is best (Sun et al., 2022).

Human evaluation in the same paper indicates that, for sampled generated knowledge, 91% was grammatical, 82% relevant, and 64% useful for answering; for generated answers, 87% was question-related and 68% useful (Sun et al., 2022). In the 2021 paper, annotation of selected knowledge showed that 83% was factually correct, 72% helpful, and 13% harmful (Liu et al., 2021). These numbers are important because they locate the main practical limitation: generated knowledge is often good, but not uniformly reliable.

A plausible implication is that the gains do not arise merely from adding more text to the prompt. Both papers report that knowledge quality and selection are decisive. In the 2021 work, random sentences generally hurt performance (Liu et al., 2021). In TSGP, generated knowledge is explicitly filtered with PMI because raw generation is noisy (Sun et al., 2022).

6. Extensions beyond commonsense question answering

Later work broadens generative knowledge prompting beyond commonsense QA while retaining the core “generate knowledge, then use it” pattern.

In prompt synthesis, “Automatic Prompt Generation via Adaptive Selection of Prompting Techniques” (Ikenoue et al., 20 Oct 2025) builds a knowledge base linking task clusters to prompting techniques and then generates prompts for new tasks from abstract task descriptions. The pipeline has two phases: knowledge base construction and prompt generation. Tasks are embedded with gemini-embedding-exp-03-07, clustered with k-means chosen by silhouette score, and each cluster is associated with 3 or 4 prompting techniques under fixed constraints such as always including Role Playing and selecting one reasoning technique (Ikenoue et al., 20 Oct 2025). The system is therefore a knowledge-driven prompt generator rather than a direct answer model.

In knowledge graph construction, “Iterative Zero-Shot LLM Prompting for Knowledge Graph Construction” (Carta et al., 2023) treats prompts as knowledge operations. Different prompts extract entities, relations, predicate descriptions, and schema-level hypernyms, progressively turning text into an explicit graph. The pipeline comprises candidate triplet extraction, entity/predicate resolution, and schema inference, all under zero-shot prompting with GPT-3.5 and without external ontologies or curated knowledge bases (Carta et al., 2023). Here, the generated intermediate content is not a QA-supporting fact but a structured knowledge artifact.

In slot filling, “Generative Zero-Shot Prompt Learning for Cross-Domain Slot Filling with Inverse Prompting” (Li et al., 2023) reformulates slot filling as seq2seq generation with T5. The model generates slot values rather than BIO tags, using a prompt of the form “what is the slot_type ?” together with all slot names. The auxiliary inverse prompting task reverses the mapping from slot value to slot type to reduce the multiple prediction problem (Li et al., 2023). This is a generative prompting method that explicitly aims to activate latent knowledge about slot semantics.

In multimodal affective computing, “Multimodal LLMs for End-to-End Affective Computing: Benchmarking and Boosting with Generative Knowledge Prompting” (Luo et al., 4 Aug 2025) uses an MLLM to generate emotion-relevant descriptions from raw audio and video, then aggregates those descriptions with the original modalities and text transcript during supervised fine-tuning. The paper frames this as a hybrid strategy that combines generative knowledge prompting with supervised fine-tuning and reports improvements such as Acc5 by 7.0% on CH-SIMS and Acc7 by 3.1% on CMU-MOSI for the optimized HumanOmni configuration (Luo et al., 4 Aug 2025).

In 3D generation, “Know3D: Prompting 3D Generation with Knowledge from Vision-LLMs” (Chen et al., 24 Mar 2026) transfers semantic knowledge from a VLM into a 3D generator through latent hidden-state injection. The VLM-diffusion bridge generates a back-view prior conditioned on a front-view image and text, and intermediate diffusion hidden states are injected into a TRELLIS2-based 3D generator via a parallel cross-attention branch with zero-initialized injection (Chen et al., 24 Mar 2026). Here, the “generated knowledge” is not text but a structured latent representation.

These extensions suggest that generative knowledge prompting has evolved from a QA-specific technique into a general design principle: a model first externalizes some intermediate semantic resource, and a later stage consumes that resource to improve control, reasoning, or structured prediction.

7. Limitations, controversies, and open directions

The core limitation reported across the literature is that generated intermediate knowledge is not guaranteed to be relevant or helpful. TSGP states explicitly that generated knowledge may be noisy because the model is not fine-tuned and that prompt design may need optimization (Sun et al., 2022). The 2021 generated knowledge prompting paper likewise shows that some selected knowledge is harmful and that hallucinated or misleading statements can degrade prediction (Liu et al., 2021).

A second recurring issue is that more generated or retrieved knowledge is not necessarily better. In the 2021 commonsense paper, performance improves up to about $[k_m \Vert q]$6 generated knowledge statements and then declines, likely because lower-quality generations introduce noise (Liu et al., 2021). In proportional analogy completion, “KnowledgePrompts” (Wijesiriwardene et al., 2024) finds that targeted knowledge is more effective than exemplar knowledge or generic structured knowledge; for GPT-3.5-Turbo, structured knowledge prompting scores around 38.29% and 38.79%, below the zero-shot result of 45.7%, while targeted knowledge prompting reaches 55.25% (Wijesiriwardene et al., 2024). This supports a strong but narrow conclusion: knowledge must be relation-specific and task-aligned, not merely present.

A third issue concerns computational and systems costs. Two-stage or multi-stage prompting requires extra sampling, filtering, or retrieval. TSGP notes computational constraints and suggests that larger models such as GPT-3 might further improve the method, while still showing that GPT-2 can already be effective (Sun et al., 2022). The original generated knowledge prompting paper reports about 200 GPU hours for inference and about \$500 for GPT-3 API generation (Liu et al., 2021).

There is also a methodological controversy around what counts as “knowledge” in these systems. In commonsense QA, the generated artifact is a natural-language statement that resembles a fact or inference (Liu et al., 2021, Sun et al., 2022). In prompt generation, the artifact is a prompting recipe (Ikenoue et al., 20 Oct 2025). In knowledge graph construction, it is a graph component or schema relation (Carta et al., 2023). In 3D generation, it is a hidden-state prior rather than language (Chen et al., 24 Mar 2026). This suggests that “knowledge” in generative knowledge prompting is best understood functionally: it is any intermediate representation intended to make a downstream decision easier, more explicit, or more controllable.

Future directions repeatedly suggested in the surveyed papers include better prompt design (Sun et al., 2022), dynamic knowledge-base updates and adaptive technique selection (Ikenoue et al., 20 Oct 2025), stronger control over hallucination (Carta et al., 2023), hybrid use with external knowledge (Sun et al., 2022), and tighter coupling between generated intermediate knowledge and alignment signals from users or task rewards (Min et al., 14 Apr 2025). A plausible implication is that the field is moving from simple two-stage prompting toward systems that jointly handle knowledge elicitation, knowledge selection, and downstream alignment.

Generative knowledge prompting therefore occupies an intermediate position between retrieval-augmented generation, chain-of-thought prompting, and structured knowledge engineering. It shares with retrieval-augmented approaches the goal of supplying missing information, but the information may be generated rather than retrieved (Liu et al., 2021). It shares with chain-of-thought the use of intermediate text, but the intermediate output is often framed as knowledge rather than a bare reasoning trace (Sun et al., 2022). It shares with knowledge engineering the goal of explicit structure, but it typically achieves that structure through prompt-mediated generation rather than schema-first symbolic design (Carta et al., 2023). Across these variants, the unifying idea remains stable: prompts can be used to expose latent model knowledge, and that exposed knowledge can materially improve downstream reasoning or generation when it is filtered, aligned, and reused effectively.