PRECEPT: Planning Resilience via Experience, Context Engineering & Probing Trajectories A Unified Framework for Test-Time Adaptation with Compositional Rule Learning and Pareto-Guided Prompt Evolution

Imagine you are hiring a very smart, but slightly forgetful and overly confident assistant to manage a complex logistics company. You give them a rulebook, but the world changes every day: ports close, new laws appear, and sometimes the rulebook itself contains outdated or even wrong advice.

Traditional AI assistants (like the ones in the paper's "baselines") try to learn by reading their own notes written in plain English. The problem? As the number of rules grows, their notes become a mess. They start misreading their own handwriting, confusing "Ship to Hamburg" with "Ship to London" because the words look similar. They get stuck in loops, repeating the same mistakes, and they can't easily combine simple rules to solve complex problems.

PRECEPT is a new framework designed to fix this. Think of it as upgrading your assistant from a "note-taker" to a "super-organized librarian with a magic rulebook."

Here is how PRECEPT works, broken down into three simple superpowers:

1. The Magic Index Card System (Deterministic Retrieval)

The Problem: Imagine trying to find a specific rule in a 1,000-page diary written in long paragraphs. If you ask, "What do I do if it's raining AND the truck is broken?", the assistant has to read the whole page, guess the meaning, and might get it wrong.

The PRECEPT Solution: Instead of a diary, PRECEPT uses a giant, perfect index card system.

Every rule is written on a card with a unique barcode (a "condition key").
When a task comes in, the assistant doesn't "read" or "interpret" the rules. It simply scans the barcode.
The Analogy: It's like using a vending machine. You don't ask the machine, "What do you think I want?" You press button "A1," and it exactly gives you the soda you asked for. No guessing, no misreading.
The Result: Even if you have thousands of rules, the assistant finds the right one instantly and with 0% error. It can also stack rules together like Lego blocks (e.g., "Safety" rules always beat "Speed" rules) without getting confused.

2. The "Trust but Verify" Detective (Conflict Resolution)

The Problem: Sometimes, your assistant has two sources of information:

The Old Manual (Static): A dusty book from 2010 that says "Always ship via Port A."
The Live News (Dynamic): A breaking news alert saying "Port A is closed."

Old assistants might get confused, trying to blend the two, or they might stubbornly stick to the old manual because it "sounds" authoritative.

The PRECEPT Solution: PRECEPT acts like a detective with a lie detector.

It treats the "Old Manual" and the "Live News" as two different witnesses.
It uses a mathematical "trust score" (Bayesian reliability). If the Old Manual says one thing and the Live News says another, and the Live News has been proven right recently, the detective immediately ignores the Old Manual.
The Analogy: Imagine a courtroom. The Old Manual is a witness who has been wrong before. The Live News is a witness who just saw the event. PRECEPT doesn't just listen to both; it weighs their credibility. If the Old Manual is lying (or just outdated), PRECEPT silences it and follows the truth.

3. The "Self-Correcting GPS" (Drift Adaptation)

The Problem: Imagine you are driving to a destination. You have a GPS that says "Turn Left." You turn left, but there's a wall. The GPS is wrong because the road changed (drift). A bad GPS would keep telling you to turn left, over and over, until you crash.

The PRECEPT Solution: PRECEPT has a smart "forget" button.

If the GPS says "Turn Left" and you hit a wall, PRECEPT doesn't just try again. It immediately deletes the "Turn Left" rule from its memory.
It marks that specific route as "Forbidden" so it never tries it again.
It then explores new paths until it finds the correct one, and then it saves the new correct rule.
The Analogy: It's like a GPS that learns from its mistakes in real-time. If a road is closed, it doesn't just say "Try again"; it permanently reroutes and updates the map so no one else gets stuck there.

The "Coach" (COMPASS)

Finally, PRECEPT has a Coach named COMPASS.

While the assistant is working, the Coach watches how well they are doing.
If the assistant is struggling with a specific type of problem, the Coach rewrites the assistant's "instruction manual" (the prompt) to make it smarter.
But the Coach is smart too: it doesn't just guess. It tests the new instructions in a simulation first. If the new instructions are better, it installs them. If not, it keeps the old ones.

Why Does This Matter?

In the real world, AI agents often fail because they try to "think" too much about simple rules. They get overwhelmed by complexity.

PRECEPT proves that structure is better than scale.

Instead of making the AI "smarter" (which is hard and expensive), PRECEPT gives it a better structure (exact rules, conflict detection, and self-correction).
The Result: In tests, PRECEPT solved complex logistics problems 40% more often on the first try than other top AI methods. It made 60% fewer mistakes, learned from its errors instantly, and could handle situations where the rules changed overnight.

In short: PRECEPT turns a chaotic, guessing game into a precise, reliable machine by giving the AI a perfect filing system, a lie detector, and a GPS that never gets stuck in a loop.

Here is a detailed technical summary of the paper PRECEPT: Planning Resilience via Experience, Context Engineering & Probing Trajectories.

1. Problem Statement

The paper addresses four fundamental limitations in current Large Language Model (LLM) agent architectures, particularly those relying on verbal reflection (e.g., Reflexion) or Reinforcement Learning (RL):

Compositional Explosion: Training on all $2^N $combinations of$ N$ atomic conditions is infeasible.
Interpretation Degradation: Verbal retrieval (natural language memory) suffers from exponential error rates as the number of conditions increases. The paper cites a theoretical model where partial-match errors reach 94.4% at $N=10$ conditions.
Drift Blindness: Agents struggle to detect when static knowledge (e.g., documentation) becomes stale or when the environment changes, often persisting with incorrect rules indefinitely.
Sample Inefficiency: RL requires prohibitive sample counts ( $\beta > 100$ ) to adapt, while verbal methods often require $\beta > 5$ , whereas practical deployment demands $\beta \le 3$ .

Existing approaches fail to jointly address these axes, leading to brittle agents that cannot compose rules reliably, detect adversarial knowledge, or adapt to environmental shifts without full retraining.

2. Methodology: The PRECEPT Framework

PRECEPT is a unified framework for test-time adaptation consisting of three tightly coupled components:

A. Deterministic Exact-Match Retrieval & Compositional Stacking

Mechanism: Instead of semantic similarity search, PRECEPT uses structured condition keys (e.g., MD5(condition_key)) to perform $O(1)$ exact-hash lookups in a rule store.
Semantic Tier Hierarchy: To handle multiple conditions, atomic precepts are organized into a hierarchy (Safety > Compliance > Preferences). When multiple rules apply, the system deterministically selects the highest-tier rule.
Compositional Generalization: By learning $N$ atomic rules, the system can theoretically resolve up to $2^N - 1$ composite scenarios without retraining, bypassing the need for LLM interpretation on the deterministic path.

B. Evo-Memory: Unified Conflict Resolution & Drift Adaptation

Inspired by the "Digital Red Queen" hypothesis, PRECEPT maintains a growing history of constraints to prevent cyclic failures.

Type I Conflict (Static vs. Dynamic): Resolves conflicts between pre-deployment static knowledge (e.g., docs) and runtime dynamic experience. It uses a six-method ensemble detector (NLI, temporal analysis, etc.) and Bayesian Source Reliability (Thompson Sampling with Beta posteriors) to weigh evidence. Dynamic experience is initially trusted more than static priors.
Type II Conflict (Rule Drift): Detects when the environment changes (e.g., valid solutions shift). It employs threshold-based rule invalidation: if a rule fails consecutively (default threshold $\theta=2$ ), its confidence decays ( $\times 0.5$ ) and it is eventually invalidated and removed from the prompt and rule store.
Deterministic Pruning: A RefineInterceptor guarantees that once a specific option is identified as a failure (Hard/Soft constraint), it is never retried ( $P(\text{repeat\_fail}) = 0$ ).

C. COMPASS: Complexity-Optimized Multi-strategy Pareto Adaptive Search

COMPASS is a dual-frequency outer loop for prompt evolution:

High-Frequency Monitor: Runs at every step ( $O(1)$ ) to check constraints, classify errors, and learn patterns.
Low-Frequency Architect: Triggered by events (e.g., goal failure), it evolves system prompts using GEPA (Generative Evolution of Prompts) enhanced with:
- ML-based Complexity Analysis: Estimates task difficulty to allocate rollouts efficiently.
- Pareto Selection: Selects prompts based on a bi-objective trade-off between Success Rate and Step Efficiency.
- MAP-Elites Diversity: Prevents "convergence collapse" by maintaining a diverse set of strategies (niche-based) rather than a single greedy strategy.
Verified Evolution: Prompts are evaluated via actual agent execution against hidden ground truth, not heuristic scoring, ensuring honest feedback.

3. Key Contributions

Compositional Rule Learning: Introduces $O(1)$ exact-match retrieval over structured keys, eliminating interpretation errors and enabling $2^N-1 $compositional coverage from$ N$ atomic rules.
Unified Conflict Resolution: A framework handling both static-dynamic conflicts (via Bayesian Thompson Sampling) and environmental drift (via threshold-based invalidation), achieving a 64x drift-resilience bound in modeled scenarios.
COMPASS Architecture: A dual-frequency adaptation layer that integrates prompt evolution with the full retrieval pipeline, using Pareto optimality and diversity maintenance to optimize for both success and efficiency.
Theoretical Guarantees: Provides closed-form bounds proving that PRECEPT avoids the exponential degradation of verbal methods and eliminates cyclic failures.

4. Experimental Results

Experiments were conducted across three domains (Logistics, Booking, Integration) with varying complexity and adversarial conditions.

First-Try Success (P1): PRECEPT achieved a +41.1 percentage point (pp) advantage over Full Reflexion and +33.3pp over ExpeL on average.
- In the "Integration" domain (high LLM prior bias), PRECEPT reached 80.0% P1 vs. 36.7% for Reflexion.
Compositional Generalization: PRECEPT achieved 100% P1 on 2-way logistics compositions and +33.3pp generalization advantage over baselines.
Continuous Learning: With minimal training ( $\beta=1$ ), PRECEPT improved by +55.0pp in logistics and +40.1pp in integration over sequential encounters, while baselines showed negligible or non-significant gains.
Drift Adaptation: When solutions changed between training and testing, PRECEPT recovered to 83.3% P1 (Logistics) and 60.7% (Integration), significantly outperforming baselines which failed to recover (e.g., Reflexion dropped to 43.8%).
Adversarial Static Knowledge: Even with poisoned static knowledge, PRECEPT's Bayesian conflict resolution allowed it to recover, maintaining a >30pp advantage over baselines.
Efficiency: PRECEPT required 61% fewer steps on average (2.32 steps vs. 5.89 for Reflexion), approaching the theoretical minimum of 2.0 steps.

5. Significance

Paradigm Shift: PRECEPT argues that reliable LLM agents require structure over scale. It resolves the tension between parametric knowledge and in-context information by routing exact tasks through deterministic retrieval and using Bayesian inference for trust, rather than relying on the LLM to "guess" the correct rule from text.
Robustness: It is the first framework to provide formal guarantees against cyclic failures and to handle both static knowledge corruption and environmental drift simultaneously.
Practicality: The framework is designed for production (MCP architecture), offering auditability (explicit rules) and high sample efficiency ( $\beta \le 3$ ), making it suitable for high-stakes domains like logistics, healthcare, and finance where black-box RL is unacceptable.
Theoretical Validation: The empirical results align with theoretical predictions, confirming that exact-match retrieval prevents the exponential degradation seen in verbal memory systems as task complexity grows.

In summary, PRECEPT demonstrates that by combining deterministic retrieval, conflict-aware memory, and Pareto-guided prompt evolution, LLM agents can achieve reliable, compositional, and adaptive performance in complex, non-stationary environments.