Contrastive Conversational Prompting (CCP)

• Contrastive Conversational Prompting (CCP) is defined as a training-free method that improves personalization by inserting negative reviews before true examples in a multi-turn dialogue.
• It employs a structured conversational dynamic where an LLM corrects an initially provided wrong review to better capture the target user's style.
• Empirical results demonstrate that CCP outperforms non-conversational baselines in key metrics like ROUGE-L, BERTScore, and user linkage, especially when hard negatives are used.

Contrastive Conversational Prompting (CCP) is a lightweight, training-free method for personalized review generation in few-shot settings, introduced in “Few-Shot and Training-Free Review Generation via Conversational Prompting” (Kusano, 25 Sep 2025). It extends Simple Conversational Prompting (SCP), which reformulates a user’s past item–review pairs as a multi-turn conversation with a LLM, by inserting negative, incorrect reviews before the true reviews and asking the model to correct them. In the reported formulation, the negative reviews are written by other users or generated by the LLM itself, and the contrastive exchange is intended to help the model distinguish the target user’s style from others’ while remaining usable when only a few user reviews are available and fine-tuning is infeasible.

1. Definition and relation to Simple Conversational Prompting

CCP is defined relative to SCP. In SCP, a user’s past item–review pairs $\{(i_1,r_1^u),\ldots,(i_n,r_n^u)\}$ are reformulated as a multi-turn conversation with the LLM. Each turn follows the same pattern: the user asks for a review of item $i_k$ based on past reviews, the assistant returns the actual review $r_k^u$, and the user then asks for the next item. The final user turn asks the LLM to generate a review for an unseen item $i_{n+1}$ (Kusano, 25 Sep 2025).

CCP extends this setup by injecting a negative review $r_k'$ before each true review. The negative review is explicitly presented as incorrect, after which the user rejects it and asks the assistant to retry; the assistant then produces the true review $r_k^u$. The reported motivation is that this contrastive feedback helps the LLM distinguish the target user’s style from others’. In the paper’s framing, SCP is the positive conversational baseline, whereas CCP is the contrastive conversational variant.

This organization places CCP within prompt engineering for low-resource personalization rather than within parameter-updating methods. A common misconception is to treat CCP as a fine-tuning procedure. The reported method is training-free: the contrast is realized entirely through prompt structure and conversational context, not through weight updates.

2. Prompt construction and conversational dynamics

The prompt structure is specified turn by turn. Let $T_{k-1}$ denote the instruction text asking for a review of item $i_k$. For the first contrastive turn, where $k=n-\ell$, the conversation begins with:

• user: $M(\text{user}, T_{k-1})$
• assistant: $i_k$0 in CCP
• user: $i_k$1
• assistant: $i_k$2
• user: $i_k$3

This pattern is repeated for $i_k$4 turns. On the last turn, there is no true review $i_k$5; instead, the LLM’s generation is collected as $i_k$6 (Kusano, 25 Sep 2025).

The paper also gives a pseudocode-style realization of the same workflow. For each historical item from $i_k$7 through $i_k$8, the prompt appends an instruction such as “Please write a review for item $i_k$9 in my style.” It then appends the negative example, the rejection message “No, that’s not me. Try again.”, the true review, and the transition “Good. Now review item $r_k^u$0.” After the loop, the final user instruction asks for a review of item $r_k^u$1, and the assistant output is returned as $r_k^u$2.

The structural difference from a conventional non-conversational prompt is therefore not merely stylistic. In the reported baseline, all past reviews are concatenated into a single user message followed by “Now review $r_k^u$3.” CCP instead serializes historical evidence into alternating user–assistant turns and introduces explicit correction signals.

3. Contrastive interpretation and formal objective

The paper presents CCP as minimizing a contrastive-style loss at each turn, while also stating that, in practice, there is no back-propagation through the LLM (Kusano, 25 Sep 2025). Let $r_k^u$4 be the LLM’s conditional distribution, let $r_k^u$5 be a similarity score such as cosine between embeddings, and let $r_k^u$6 be a temperature. When the LLM proposes review $r_k^u$7 for item $r_k^u$8, the objective is to make $r_k^u$9 close to the true review $i_{n+1}$0 and far from the negative review $i_{n+1}$1.

The turn-level loss is given as an InfoNCE-style objective:

$i_{n+1}$2

Summing over turns gives the overall CCP loss:

$i_{n+1}$3

The paper explicitly qualifies this formalization: the loss is a way to view the method, not an optimization actually applied to model parameters. Instead, the conversation “guides” the LLM to implicitly follow this objective. This distinction is important for interpretation. CCP is not presented as a new training criterion for LLM adaptation, but as a prompt-level mechanism whose behavior can be described in contrastive terms.

A plausible implication is that the formalism functions chiefly as an explanatory bridge between conversational prompting and contrastive representation learning, rather than as an implementational requirement. That implication follows from the paper’s explicit statement that there is no back-propagation.

4. Negative examples and variant design

Negative-example selection is central to CCP. The summary distinguishes negatives drawn from real users and negatives generated by the LLM itself (Kusano, 25 Sep 2025). When negatives are taken from real users, several variants are defined by the retrieval criterion:

Variant	Negative source	Selection rule
CCP(R)	Other-user review	Highest ROUGE-L overlap to $i_{n+1}$4
CCP(B)	Other-user review	Highest BERTScore
CCP(R)$i_{n+1}$5 / CCP(B)$i_{n+1}$6	Other-user review	Lowest-scoring review
CCP(G)	LLM-generated review	Initial single-turn generation

For real-user negatives, CCP(R) chooses the other-user review $i_{n+1}$7 with highest ROUGE-L overlap to the true review $i_{n+1}$8, yielding a “hard” negative. CCP(B) chooses the one with highest BERTScore. The easy-negative variants, CCP(R)$i_{n+1}$9 and CCP(B)$r_k'$0, choose the lowest-scoring review. For LLM-generated negatives, CCP(G) first uses an initial single-turn prompt to generate $r_k'$1 and then treats that generated review as the negative.

The reported finding is that hard negatives, defined as high semantic or textual similarity negatives, yielded the biggest gains, whereas low-quality negatives had minimal effect. The practical interpretation offered by the paper is correspondingly selective: when similar reviews can be collected, BERTScore-based hard negatives are recommended; when other users’ reviews are inaccessible, generated negatives remain possible at the cost of extra API calls.

This suggests that CCP’s benefit depends not only on conversational formatting but also on the informativeness of the contrast. The method is therefore sensitive to the quality of the negative channel, even though it remains training-free.

5. Experimental setting and comparative performance

The reported experiments use eight Amazon product categories: Movies, Music, Books, Kindle, Groceries, Games, Sports, and Electronics. Each category contains 200 users, and reviews have length 20–300 tokens. History length is varied over $r_k'$2, with conversation turns $r_k'$3 typically set to $r_k'$4. The LLMs evaluated are GPT-4.1-mini, GPT-4.1, o4-mini, llama3.3-70b, and claude-sonnet-4, all at temperature $r_k'$5 (Kusano, 25 Sep 2025).

Evaluation is carried out with both text-based and application-oriented metrics. The text-overlap metrics are ROUGE-L and BERTScore. User-identity linkage is measured with Hit@5 and MRR by ranking the generated review among other-user reviews. Sentiment alignment is assessed using KL divergence of group-level histograms and user-level Weighted-F1 and Macro-F1 from a RoBERTa classifier.

In the comparative analysis, the non-conversational prompt is described as a baseline consisting of a single user message concatenating all $r_k'$6 past reviews and “Now review $r_k'$7.” Its reported result is ROUGE-L $r_k'$8 and BERTScore $r_k'$9, described as indistinguishable from random-user reviews. SCP improves this to ROUGE-L $r_k^u$0–$r_k^u$1 and BERTScore $r_k^u$2. CCP further improves to ROUGE-L $r_k^u$3–$r_k^u$4 and BERTScore $r_k^u$5–$r_k^u$6 (Kusano, 25 Sep 2025).

A representative configuration is explicitly reported for $r_k^u$7, $r_k^u$8, and GPT-4.1-mini:

Method	ROUGE	BERTScore
Baseline	0.133	0.848
SCP	0.155	0.858
CCP(B)	0.161	0.861

For user linkage, Hit@5 is reported as approximately $r_k^u$9 for the baseline, approximately $T_{k-1}$0 for SCP, and approximately $T_{k-1}$1 for CCP(B). For sentiment KL divergence, the reported values are approximately $T_{k-1}$2 for the baseline, approximately $T_{k-1}$3 for SCP, and approximately $T_{k-1}$4 for CCP(B), where lower values indicate better matching.

The broad empirical conclusion is that both conversational variants produce reviews much closer to those of the target user, even when each user has only two reviews, and that CCP brings further improvements when high-quality negative examples are available.

6. Operational guidance, limitations, and interpretation

The paper gives several practical guidelines. When no external reviews are available, SCP is reported to give large gains while costing only one prompt round per history item. If similar reviews by the same user or by others on the same item can be collected, the recommendation is to select hard negatives via BERTScore, corresponding to CCP(B). If other users’ reviews are unavailable, negatives can be generated via the LLM itself, yielding CCP(G), though this requires extra API calls. The paper also states that using $T_{k-1}$5 conversation turns provides maximal benefit and that inserting negatives at every turn yields the largest improvement (Kusano, 25 Sep 2025).

The main limitation identified in the summary is inference cost. CCP(G) can be 3–4$T_{k-1}$6 more expensive than SCP, so generating many items at scale may require balancing cost against quality. The summary further notes that a cheaper LLM plus SCP can outperform a more expensive model plus the non-conversational baseline. That point does not claim universal superiority for SCP over stronger models; rather, it specifies a cost–quality trade-off under the reported prompting regimes.

Another interpretive boundary concerns the role of the formal loss. Because the method does not back-propagate through the LLM, the contrastive objective is explanatory rather than directly optimized. A related misconception would be to assume that any negative example is equally useful. The reported results do not support that: low-quality negatives had minimal effect, whereas hard negatives yielded the biggest gains.

Taken together, the reported evidence positions CCP as a training-free technique for cold-start personalized review generation under few-shot constraints. The core claim is not that conversational prompting replaces learned personalization in all settings, but that reformulating limited user history as a corrective multi-turn dialogue can improve stylistic fidelity and downstream matching relative to a standard single-turn prompt.