Answer Engineering: Local Trajectory Editing for Protocol-Constrained Decision Making in Large Language Models

Technical Summary: Answer Engineering

Problem Statement
Large Language Models (LLMs) often generate confident but incorrect answers in domains requiring strict adherence to procedural protocols, such as clinical medicine. While step-by-step reasoning techniques (e.g., Chain-of-Thought) improve structure and transparency, they frequently fail to prevent errors in protocol-constrained settings. Instead of eliminating failures, reasoning often shifts the distribution of errors. Empirical studies indicate that models may generate explanations that function as post-hoc rationalizations rather than faithful derivations of the final answer. Consequently, models may commit to a decision early and rationalize it, or generate intermediate steps that are inconsistent with domain rules. In regulated environments, an answer may appear plausible yet be operationally invalid if it omits mandatory procedural checks or distinctions. Existing approaches, including prompting, constrained decoding, and post-hoc validation, often address these issues indirectly, globally, or only after generation is complete, failing to intervene within the evolving reasoning trajectory itself.

Methodology: Answer Engineering
This paper proposes Answer Engineering, a deterministic runtime and authoring layer that performs localized trajectory intervention during autoregressive generation. Unlike methods that require model retraining or global search, Answer Engineering operates as a control surface over standard decoding, enforcing protocol compliance through rule-based edits to the generated reasoning trace.

The core mechanism relies on the property of prefix-conditioned decoding: in an autoregressive transformer, the distribution of future tokens depends solely on the visible prefix. Therefore, local edits to the reasoning trajectory redefine the context for subsequent generation without altering the model parameters.

The framework implements three classes of intervention, triggered by declarative rules authored by domain experts:

1. Editing the Past (Retroactive Span Replacement): When a trigger indicates a violation (e.g., an incorrect inference), the system replaces the offending span with a valid candidate selected based on model likelihood.
2. Selecting Among Recent Continuations (Local Rollback): Upon detecting a trigger, the system rolls back to the start of the edit scope, generates multiple candidate continuations, and selects the highest-scoring valid one according to protocol rules.
3. Forcing the Future (Insertion): If a protocol requires a specific statement (e.g., an urgency warning), the system appends a predefined valid candidate to the current prefix and advances the generation through these forced tokens.

The runtime operates as a control loop: it monitors the trajectory for rule triggers, proposes candidate edits, scores them using the model's own likelihood function, and applies the selected intervention before resuming decoding. This process involves rebuilding the KV cache from the earliest modified token to ensure the generation continues from the corrected state.

Key Contributions
The paper makes three primary contributions:

1. A Decoding-Time Runtime Mechanism: A framework for enforcing protocol compliance through localized trajectory intervention during generation, eliminating the need for model retraining or global search.
2. A Rule-Based Control Abstraction: A method for specifying domain-specific procedural constraints as interpretable trajectory-editing operations, allowing domain experts to define permissible edits without modifying model code.
3. Empirical Evaluation: A demonstration that step-by-step reasoning alone may reduce protocol compliance in structured decision tasks, and that localized trajectory intervention can substantially improve compliant outcomes.

Experimental Results
The approach was evaluated on a clinical benchmark involving Sudden Sensorineural Hearing Loss (SSNHL), requiring differential diagnosis between SSNHL and conductive hearing loss based on Weber and Rinne tuning-fork findings. The evaluation used a deterministic dataset of 2,000 synthetic cases (1,000 SSNHL, 1,000 conductive) and the OpenMeditron/Meditron3-8B model.

Key findings include:

• Reasoning Baseline Failure: Under a step-by-step reasoning baseline, compliant outcomes for the target SSNHL condition dropped from 54.5% (unguided) to 25.1%, while performance on the contrastive conductive condition increased from 1.6% to 58.9%. This indicates that reasoning alone shifted errors rather than preventing them.
• Trajectory Editing Success: Applying the full trajectory editing stack (combining local validation, replacement, and forced insertion) increased compliant outcomes for the target SSNHL condition to 83.5% and protocol adherence for the contrastive conductive condition to 77.9%.
• Balanced Accuracy: The balanced accuracy across both conditions improved from 42.0% (reasoning baseline) to 80.7% with trajectory editing.
• Runtime Overhead: The intervention incurred a slowdown of 2.78× for SSNHL cases and 1.33× for conductive cases relative to the reasoning baseline.
• Failure Analysis: The authors note that while the method significantly improves outcomes, it does not guarantee uniform reasoning stability. In some cases, the model exhibits "commitment persistence," repeatedly attempting premature diagnostic closure despite repeated local repairs.

Significance and Claims
The paper claims that Answer Engineering demonstrates that reliability in protocol-governed reasoning is not solely a property of model training but can be treated as an operational property shaped by runtime control. The results suggest that many failures in protocol-constrained generation are structural violations (e.g., omitted steps, local inconsistencies) that can be corrected through domain-specific protocols without retraining.

The authors position this work as a feasibility study under structured conditions. They explicitly state that the evaluation isolates a well-defined diagnostic boundary to make reasoning failures observable and correctable. They do not claim this as a general solution for all reasoning reliability issues, acknowledging that real-world workflows involve interacting protocols, partial observability, and evolving guidelines. The significance lies in providing a systems-level perspective where protocol adherence is engineered through explicit control of generation behavior, offering a path to improve compliance in regulated domains where standard reasoning techniques have proven insufficient.