RECAP: Regression Evaluation for Continual Adaptation of Prompts

Abstract: Production agentic systems routinely face evolving constraints and must comply from the very next interaction. Scenarios like a tool-call notification changing a compliance threshold or a policy update adding disclosure requirements fit this criteria, having close to no room for errors in production. This proactive adaptation setting is common in deployment, but absent from current benchmarks, which assume either static constraint sets or reactive protocols with evaluation feedback. We introduce RECAP, a benchmark that measures continual-learning phenomena (forgetting, regression, forward transfer) at the constraint level under a strictly proactive adapt-then-test protocol: prompt optimization methods receive only the constraint specification and must generalize before seeing any test data. Evaluating six methods across four LLMs and three schedules with evolving constraints, we find that these methods show no significant improvement in performance, even after incurring a higher latency. These methods, designed for offline or reactive settings, are inadequate for the proactive paradigm. Our work emphasizes the growing need for designing proactive prompt adaptation methods, where the models must remain robust to evolving needs in deployment.

• The paper shows that no adaptation method outperforms a baseline that simply appends current active constraints.
• It presents the RECAP framework which evaluates continual learning phenomena such as forgetting, collateral damage, and unlearning fidelity.
• Experiments demonstrate that model scale—rather than adaptation strategy—drives performance in handling dynamic operational constraints.

RECAP: Toward Proactive Prompt Adaptation under Continually-Evolving Constraints

Problem Formulation and Motivation

Agentic LLM systems in production environments routinely face dynamically shifting operational constraints due to policy updates, user preference changes, or evolving tool requirements. These constraint modifications demand strict adherence from the very next interaction, with essentially zero tolerance for adaptation lags or compliance errors. Existing benchmarks—typically static or reliant upon test-time feedback (reactive protocols)—fail to capture the proactive adaptation regime where an agent must satisfy unseen, updated constraints immediately and without feedback. The challenge is to support continual learning at the prompt level, when underlying model weights are frozen and only prompt augmentation is permitted.

The authors introduce the RECAP benchmark to bridge this gap, focusing on proactive, constraint-level continual adaptation. Here, the prompt adaptation method observes only the new or edited constraint specification before deployment and is evaluated against all active (including edited or deleted) constraints on held-out test data, with no opportunity for reactive correction using explicit feedback. This setting operationalizes the adapt-then-test protocol, highlighting the unique demands of production LLM deployment not addressed by prior CL or prompt optimization literature.

Benchmark Design and Methodological Framework

RECAP decomposes static instruction-following benchmarks into temporal evaluation streams by transforming static instruction/constraint pairs into sequential schedules of constraint add, edit, and delete operations. Each step of a schedule initiates a single modification and evaluates continual learning phenomena—including forgetting, regression, and forward transfer—by maintaining shadow evaluation on obsolete (edited or deleted) constraints alongside the current active set.

For measurement, RECAP introduces a decomposed metric suite:

• Mean Constraint Satisfaction ($\overline{\text{sat}}$): Average compliance across all constraints and schedule steps.
• Peak Forgetting: Maximum drop in satisfaction for a constraint from its prior best value.
• Collateral Damage: Degradation in non-targeted constraints following an operation.
• Edit Switch and Unlearning Fidelity: Quantifying prompt uptake of changes and efficacy of unlearning removed constraints, respectively.

The evaluation protocol enforces proactivity—adaptation methods see only constraint specifications (not test data or evaluation results) before being evaluated. Synthetic self-play (LLM-generated training data) is permitted internally, but no history or feedback about past evaluation outcomes is provided. A no-adaptation baseline (Base LLM) appends only the current active constraints to user prompts at inference, providing a strong reference for comparison.

Six adaptation methods spanning few-shot ICL, memory-based (Dynamic Cheatsheet), and optimization-based (ACE, GEPA, MIPROv2) paradigms are evaluated across four foundation LLMs (Llama-3.1-8B/70B, GPT-OSS-20B/120B) on three constraint evolution schedules.

Empirical Findings and Analysis

Across all 72 tested conditions, no adaptation method outperforms the no-adaptation baseline on any evaluation metric, even when incurring up to 1.7x latency and substantial increases in prompt context length. Notably, adaptation is actively detrimental on GPT-OSS models, with mean constraint satisfaction reductions up to $-0.176$. Llama models are robust to proactive adaptation but show negligible improvement over the baseline despite increased computational overhead. These trends persist across schedules with both deterministic and LLM-judged evaluations.

Several structural limitations of current adaptation approaches are revealed:

• Self-play optimization fails to generalize: High synthetic constraint satisfaction during self-play adaptation does not translate to real evaluation data, particularly as operations compound.
• Context bloat harms retention: As adaptation mechanisms append more rules or demonstrations, critical constraint signals are diluted, and out-of-date or hallucinated instructions exacerbate confusion, triggering refusal cascades or stale directive following.
• Information architectural gap: When adaptation is allowed access to all active constraints (not just the latest δ), performance recovers but does not surpass the baseline, and at the cost of massive context expansion.
• Failure to unlearn: Even when constraints are deleted, semantic relatedness between topic and keyword constraints prevents high-fidelity unlearning, with deleted constraints continuing to influence outputs in a significant fraction of cases.

Detailed behavioral analysis further identifies six primary failure modes, including specification lock (unchanging, obsolete prompt components), refusal cascades due to internally perceived contradictions, prefix contamination from synthetic demonstrations or evolved prompts, and inert memory growth where adaptation mechanisms provide no marginal utility.

Crucially, model scale, not adaptation strategy, dominates performance. Larger models perform better at constraint following regardless of adaptation, and smaller models are harmed most by adaptation artifacts.

Implications and Future Directions

RECAP demonstrates that, in a realistic, proactive adaptation scenario, prompt optimization methods designed for static or reactive continual learning regimes do not transfer. The optimal strategy under current architectures is simply to append the active constraint set to the prompt, with any additional meta-prompt engineering, self-play, or adaptive memory conferring no benefit and often causing harm. For production scenarios where new requirements must be met before exposure to any real test data, reliance on conventional prompt adaptation or CL strategies risks both regression and costly inefficiency.

This establishes a clear agenda for future research: architectures and adaptation mechanisms must be explicitly designed for proactive, prompt-level continual adaptation, with careful attention to context bloat, robust unlearning, and constraint specificity. Hybrid fine-tuning approaches, prompt-centric continual learning with better context management, and novel representations for constraint specifications may be necessary. Furthermore, theoretical advances are needed to understand the limits of frozen-model prompt adaptation under ever-evolving constraint sets, especially in safety-critical and compliance-driven real-world deployments.

Conclusion

RECAP provides a rigorous framework for evaluating proactive prompt adaptation under evolving constraints, revealing that current adaptation variants are structurally inadequate for deployment in production agentic systems. The findings articulate both the practical urgency and the technical challenges of proactive prompt adaptation, and underscore the need for principled research into regression-resistant, efficient, and context-robust prompt learning paradigms suitable for continually evolving operational environments.

---

Reference:

"RECAP: Regression Evaluation for Continual Adaptation of Prompts" (2606.06698)