ContrAstive Prompt Orchestration (CAPO)

• CAPO is a formal learning paradigm that employs contrastive objectives on diverse prompt variants for optimized downstream performance.
• It utilizes discrete ranking and adaptive aggregation methods, integrating contrastive losses to enhance language model optimization and policy transfer.
• Empirical results reveal significant gains in prompt quality, safety alignment, and few-shot learning across NLP and embodied AI tasks.

ContrAstive Prompt Orchestration (CAPO) is a formal learning paradigm and algorithmic family for leveraging prompt-level contrast—sometimes with dynamic orchestration mechanisms—to optimize downstream model behavior. Central to CAPO is the use of contrastive objectives over prompt variants, system prompts, or prompt-derived prototypes, and the explicit orchestration of these prompt forms (including dynamic aggregation). The approach is applied in diverse areas including LLM prompt optimization, few-shot learning, safety-aligned generation, unsupervised embedding construction, and visuomotor policy transfer. Key instantiations and theoretical formulations derive from large-scale empirical studies across NLP and embodied AI domains (Lee et al., 2 Sep 2025, Zhang et al., 1 Feb 2026, Zhao et al., 2024, Zeng et al., 2022, Jian et al., 2022).

1. Formal Definition and Framework Variants

CAPO is characterized by the following elements:

• Contrast Source: Contrasting prompt structures, exemplars, or learned soft prompts with respect to quality, task relevance, or domain specificity.
• Orchestration Mechanism: Either discrete selection (e.g., ranking and partitioning) or adaptive aggregation (via attention or optimization) over a set of learned prompts.
• Contrastive Objective: Explicit use of contrastive loss functions (InfoNCE and its variants), supervised or unsupervised, to align model outputs with desired invariances or discriminative boundaries.
• Application Domain: From discrete prompt optimization for LLMs to learnable prompt pools in vision-language policy learning.

A general CAPO mapping is:

$\mathrm{CAPO}: (q, \mathcal{P}) \mapsto p^*,$

where $q$ is a query, $\mathcal{P}$ a pool of candidate prompts (structured, learned, or retrieved), and $p^*$ the orchestrated prompt yielding optimal downstream performance under a designated metric.

Discrete retrieval-augmented CAPO (Lee et al., 2 Sep 2025):

• For LLM prompt optimization, $\mathcal{R}(q)$ retrieves $k$ scored prompts $p_i$ (e.g., from HelpSteer2), partitioned or ranked by metrics $$M = \{\text{help}, \text{corr}, \text{coh}, \text{comp}, \text{verb}\}$$.
• Contrastive reasoning operators $\Phi_{\mathrm{CR}}$ and $\Psi_{\mathrm{CR}}$ generate reasoning-augmented instructions for $f_\theta$, leading to a synthesized optimized prompt.

Continuous and adaptive CAPO (Zhang et al., 1 Feb 2026):

• For cross-embodiment visuomotor policy, a learnable pool $\mathbf{P}$ of prompts is established, each disentangling a distinct domain factor (e.g., lighting, FOV).
• Adaptive orchestration $\mathcal{G}_{\mathrm{attn}}$ dynamically aggregates these via attention-weighted fusion, conditioning prompt mixing on current observations.

2. Retrieval-Augmented and Tiered Contrastive Prompt Optimization

In the automatic prompt optimization setting (Lee et al., 2 Sep 2025), CAPO proceeds via:

1. Retrieval: Given a query $q$, $\mathcal{R}(q)$ retrieves $k$ prompts annotated for $M$. BM25 scoring for retrieval is used:

$$\mathrm{score}_{\mathrm{BM25}}(q,p) = \sum_{t \in q} \text{IDF}(t)\; \frac{\text{tf}(t,p)(k_1 + 1)}{\text{tf}(t,p) + k_1(1-b + b\,|p|/\text{avgLen})}$$

1. Prompt Partitioning and Contrast Formation:
   • Prompts are sorted by average metric score $\overline{s}(p_i)$.
   • Disjoint sets are formed: $P^H$ (top), $P^M$ (middle), $P^L$ (bottom).
2. Contrastive Reasoning Instruction:
   • Inputs to $f_\theta$ use templates that reflect on strengths (from $P^H$), weaknesses (from $P^L$), and stable attributes (from $P^M$).
3. Objective:
   • Margin-based notional objectives guide contrastive distance between embeddings for high- and low-quality prompts, albeit with black-box $f_\theta$.

Alternative "metric-wise" contrast isolates best-per-metric exemplars $P^m$, instructing $f_\theta$ to synthesize a composite prompt integrating strengths across all dimensions.

3. Hybrid Contrastive Prompt Pool Learning and Dynamic Orchestration

For cross-embodiment visuomotor adaptation (Zhang et al., 1 Feb 2026), CAPO incorporates:

• Pool Construction: $K$ learnable continuous prompts, each trained with a hybrid of:
  • Visual InfoNCE: enforcing invariance to lighting/appearance variations
  • Temporal action-based BYOL: aligning across embodiment/trajectory sequences
  • Text-to-vision alignment: semantic grounding via CLIP-based contrastive loss
• Adaptive Orchestration Mechanism: Given an observation $o_t$, embeddings $z_t^k = \Phi(o_t, p^k)$ for each prompt $p^k$ are attention-weighted:

$$\alpha_k = \frac{\exp(s_k^a \cdot s_k^c)}{\sum_j\exp(s_j^a \cdot s_j^c)}$$

with $s_k^a$ (learnable MLP score) and $s_k^c$ (cosine similarity with unprompted $z_t^v$).

• Fused Representation: The final feature is $$z_t^f = z_t^v + z_t^t + \sum_{k=1}^K \alpha_k\,z_t^k$$, input to a policy optimized by PPO.

4. Contrastive Orchestration in Safety Alignment and Decoding

In safe LLM alignment, Adversarial Contrastive Decoding (ACD) (Zhao et al., 2024) instantiates a CAPO framework with:

• Dual Opposite Prompt Optimization:
  • Learning two soft prompts—Safeguarding Prompt (SP) and Adversarial Prompt (AP)—via prompt-tuning on an anchor set distinguishing “refused” vs. “accepted” outputs in harmful/benign instruction cases.
  • Separate losses are applied to reinforce or discourage harmful completions.
• Contrastive Decoding:
  • At inference, logits under SP and AP are combined by $l_{\mathrm{ACD}} = l_S - \alpha l_A$, directly subtracting unsafe responses as evidenced by AP.
  • This orchestration consistently boosts harmlessness (HLR) across models by over 20 percentage points, while maintaining performance on regular tasks.
• Comparison with Other Methods:
  • ACD requires no second model and auto-learns both prompt legs, outperforming non-learned or single-template contrastive approaches.

5. CAPO in Few-Shot and Unsupervised Representation Learning

Few-shot prompt-based learning (Jian et al., 2022):

• CAPO generates multiple prompt+demonstation "views" per example, differing in template or context.
• A supervised contrastive loss $\mathcal{L}_{\text{con}}$ clusters same-class prompt views and repels cross-class ones, supplementing the masked-LM loss.
• Results show +2–6 percentage points gains in accuracy/F1 over strong prompt-only and retrieval-augmented baselines across 15 tasks.

Unsupervised sentence embedding (Zeng et al., 2022):

• ConPVP constructs prompt-derived virtual semantic prototypes, with each instance paired to both positive and negative prompt-based sequences.
• A prototypical InfoNCE loss pulls anchor sentence embedding $v_i$ towards its positive prototype $p_i^+$ and away from its negative prototype $p_i^-$ plus all other batch prototypes.
• Empirical results show consistent improvements in STS tasks (e.g., +2.6 Spearman’s $\rho$ over SimCSE) and text clustering accuracy.

6. Experimental Results and Empirical Findings

Prompt Optimization (LLMs, (Lee et al., 2 Sep 2025))

Model	Help	Corr	Coh	Comp	Verb	Avg
GPT-4o Direct	0.366	0.435	0.767	0.405	0.664	0.527
CAPO-Tiered	0.525	0.607	0.882	0.447	0.717	0.636
CAPO-Metric	0.516	0.596	0.876	0.432	0.678	0.620

• Ablations indicate that omitting contrastive reasoning degrades performance by 8–12%.
• k=10 retrieval is optimal; larger $k$ leads to noise dilution.

Visuomotor Policy Transfer (Zhang et al., 1 Feb 2026)

Approach	SR↑	SPL↑	NE↓	EL↓
CURL	52.0±1.9	0.32±0.07	0.48±0.08	32±6
CAPO	97.9±1.2	0.66±0.04	0.02±0.01	18±3

• Ablations confirm the complementary necessity of visual, temporal-action, and text contrastive objectives.
• CAPO exhibits superior zero-shot generalization across domains and embodiment changes.

Safety Decoding (Zhao et al., 2024)

• HLR (Harmless Rate) is improved from 71.4% (Base) to 92.4% (ACD CAPO), with negligible cost to general win/truthful rates and halved jailbreak attack success rates.

7. Limitations, Open Questions, and Future Directions

Limitations across CAPO studies include:

• Dependence on annotated prompt corpora (e.g., HelpSteer2) or domain-aligned anchor sets (Lee et al., 2 Sep 2025, Zhao et al., 2024).
• Generalization to multi-turn dialogue and unannotated domains is largely unexplored.
• In adaptive orchestration settings, prompt pool size and length trade-offs exist (poor performance with excessive redundancy or overfitting) (Zhang et al., 1 Feb 2026).
• Retrieval quality (BM25 vs. neural) and model-agnostic orchestration merit further research.

Suggested future work encompasses:

• Dynamic, user-driven metric weighting in orchestration.
• Integration of human-in-the-loop for iterative prompt refinement.
• Orchestration over chain-of-thought or multi-step reasoning traces.
• Extension of CAPO-inspired approaches to continuous prompt spaces across broader multimodal and multilingual domains.

• "Better by Comparison: Retrieval-Augmented Contrastive Reasoning for Automatic Prompt Optimization" (Lee et al., 2 Sep 2025)
• "Learning Adaptive Cross-Embodiment Visuomotor Policy with Contrastive Prompt Orchestration" (Zhang et al., 1 Feb 2026)
• "Adversarial Contrastive Decoding: Boosting Safety Alignment of LLMs via Opposite Prompt Optimization" (Zhao et al., 2024)
• "Contrastive Learning with Prompt-derived Virtual Semantic Prototypes for Unsupervised Sentence Embedding" (Zeng et al., 2022)
• "Contrastive Learning for Prompt-Based Few-Shot Language Learners" (Jian et al., 2022)