soft prompt tuning

To address the degradation of prior prompt information and reduced generalization caused by layer-wise replacement of deep continuous prompts in vision-language models, this paper proposes Modular Prompt Learning (MPL). MPL introduces, for the first time, a modular prompt architecture that is both accumulative and reusable: it explicitly preserves and reuses prompts from preceding Transformer layers, departing from conventional single-layer overwrite paradigms. By employing parameter tying and hierarchical prompt retention, MPL optimizes only lightweight prompt parameters while keeping the backbone frozen. Evaluated on 11 benchmark datasets, MPL achieves an average 0.7% improvement in base-to-novel class generalization, with a notable 10.7% gain on EuroSAT—outperforming existing prompt learning methods. It effectively mitigates prompt decay and enhances cross-dataset and zero-shot transfer stability.

This work addresses the overfitting of base classes and degraded generalization to novel classes in few-shot scenarios when fine-tuning only class-specific prompts in CLIP. To mitigate this, the authors propose Concept-Constrained Prompt Learning (CCPL), a lightweight regularization approach that operates under frozen CLIP encoders. CCPL aligns learnable prompts with predefined concept-level textual prototypes through shared context tokens, incorporates concept dropout for regularization, and employs a controllable logit fusion mechanism. Experiments demonstrate that CCPL improves the harmonic mean of base and novel class accuracy by 0.6% on DTD and 2.9% on EuroSAT, while incurring only a marginal 0.1% drop on OxfordPets, thereby validating its effectiveness and delineating its applicability boundaries in enhancing cross-category generalization.

This work addresses the challenge of quantitatively evaluating latent capabilities—particularly harmful behaviors—in language models. We propose **conditional distance** as a unified metric, formalizing soft prompt optimization as the minimal semantic perturbation required to activate a target behavior—introducing the first principled formulation of *conditional behavioral reachability*. We design a novel **generalized conditional soft prompting** framework enabling consistent, cross-task (e.g., NLP, chess, grid pathfinding) and cross-domain evaluation. Integrating gradient-driven embedding optimization, conditional behavioral modeling, and an automated pipeline, our approach yields interpretable, comparable, and scalable latent capability probing. Experiments demonstrate that our method effectively quantifies the difficulty of behavior activation, providing red-teaming assessments with actionable, quantitative feedback.

This work investigates the intrinsic difficulty of prompt engineering for large language models (LLMs): why effective prompts remain hard to discover and interpret reliably—even when deployed on near-optimal pretrained sequence predictors. We formalize prompt difficulty from two complementary perspectives—statistical learning theory and empirical realizability—treating prompts as conditional modulators of the pretrained distribution. Using systematic exhaustive search over binary sequence prediction tasks, explicit modeling of the pretrained distribution, and analysis of neural predictor performance bounds, we demonstrate that optimal prompts exhibit counterintuitive structural properties and depend critically on unobservable characteristics of the pretrained data distribution. Intuitively designed or task-sample-based prompts are provably suboptimal. These findings challenge prevailing prompt design paradigms and establish a novel theoretical foundation for studying prompt interpretability, optimization, and principled evaluation benchmarks.

Current prompt engineering research lacks systematic taxonomies and comparable evaluation protocols. Method: This paper introduces the first comprehensive taxonomy spanning large language models (LLMs) and multimodal models, categorizing over one hundred prompt techniques by application scenario and uniformly specifying their supported models, benchmark datasets, and boundary conditions. We integrate bibliometric analysis, cross-model/cross-dataset empirical comparison, methodological abstraction, and taxonomy construction techniques; further proposing a standardized evaluation framework and an interactive knowledge graph to clarify strengths, limitations, and open challenges of each technique. Contribution/Results: We deliver a structured technical survey, a complete classification table, and reusable evaluation dimensions—establishing the first authoritative benchmark and research navigation toolkit for prompt engineering.

This study addresses the instability and lack of reliability analysis in large language model outputs caused by opaque and redundant prompt structures. We propose a few-shot prompt minimization framework from a black-box perspective. By integrating black-box optimization-based ablation experiments, character-level statistical analysis, and logical identifier preservation strategies, the method distills prompts to their causally essential subsets while revealing divergent model preferences when functioning as universal encoders versus decoders. Experiments demonstrate that the framework achieves an average 65.3% reduction in character count while fully preserving propositional output fidelity. These findings confirm that models preferentially rely on logical identifiers over natural language descriptions, which can be safely discarded. This work establishes a novel paradigm for efficient prompt compression and structural interpretation.

This paper challenges the theoretical foundations and explanatory power of “text gradient”-based automated prompt optimization methods, which metaphorically equate discrete text updates with continuous, differentiable gradient descent. Method: Through systematic LLM prompt fine-tuning experiments, multi-task comparative analysis, ablation studies, and behavioral attribution, we rigorously examine whether these methods operate as genuine gradient-based optimizers. Contribution/Results: We demonstrate that performance gains are not attributable to gradient update logic; instead, “text gradients” function merely as empirical heuristics without theoretical grounding in differentiable optimization. First, we formally establish their non-gradient nature. Second, we propose a novel conceptual framework for prompt optimization explicitly tailored to discrete text spaces. Third, we advocate shifting prompt engineering from analogical transfer (e.g., borrowing optimization metaphors from continuous domains) toward intrinsic, ontology-aware modeling. These findings call for a fundamental methodological rethinking of prompt optimization.

This study addresses the computational overhead, inference latency, and accuracy degradation caused by prompt redundancy in large language models (LLMs) by proposing a novel prompt minimization paradigm. Methodologically, we construct an LLM-based multi-version prompt optimization and evaluation framework that employs three strategies to identify the most concise, high-density inputs. This work reveals substantial redundancy within the input space, establishes new evaluation criteria for minimalist prompts, and redefines the theoretical boundaries of efficient prompt engineering. Experimental results demonstrate that minimal prompts maintain output fidelity comparable to their longer counterparts while significantly reducing inference costs and enhancing overall system efficiency.

This work addresses the challenge of automatically discovering interpretable and discriminative global features from unstructured text by proposing a dataset-level prompt optimization approach. It extends prompt learning beyond the instance level to the dataset level for the first time, employing a multi-agent collaborative framework that iteratively generates feature definitions, extracts feature values, and jointly optimizes a shared prompt based on feedback from both downstream classification performance and interpretability. Experimental results demonstrate that the method automatically produces high-quality, human-understandable feature sets across multiple text classification tasks, significantly enhancing model performance and confirming its effectiveness and generalizability.

Existing prompt optimization methods treat prompts as monolithic units, making it difficult to localize errors, preserve critical instructions, or control prompt inflation—limitations that particularly hinder the reasoning performance of small open-source models. This work proposes Modular Prompt Optimization (MPO), a novel framework that introduces segment-wise local optimization based on fixed semantic blocks such as system roles and task descriptions. MPO leverages a critic model to generate text gradients at the segment level, enabling independent refinement of each module followed by deduplicated fusion, thereby achieving interpretable, robust, and efficient optimization while preserving the original prompt structure. Evaluated on the ARC-Challenge and MMLU benchmarks, MPO substantially outperforms both raw prompts and the TextGrad baseline, significantly boosting the reasoning accuracy of LLaMA-3-8B-Instruct and Mistral-7B-Instruct.