Technical Summary: In-Place Feedback: Reliable Refinement for Multi-Turn Expert-LLM Collaboration

Problem Statement

Large Language Models (LLMs) are increasingly deployed as collaborative assistants in expert workflows (e.g., software engineering, legal drafting, clinical documentation). However, these tasks often involve complex reasoning where mistakes can cascade. While experts typically review outputs and provide fine-grained feedback to refine the model's response, prior work indicates that LLMs frequently fail to reliably integrate this feedback in multi-turn interactions.

The core issue is "context pollution." As dialogue turns accumulate, the model must integrate the latest correction while preserving earlier feedback and leaving unrelated content untouched. When this integration fails, the model may:

1. Corrupt previously correct content.
2. Ignore the new feedback and reproduce prior mistakes.
3. Incorporate feedback locally but introduce new errors in subsequent reasoning.

These failures force experts to re-check outputs and re-state corrections, undermining productivity and causing fatigue.

Methodology: In-Place Feedback

The authors propose in-place feedback, a new interaction paradigm where the user directly edits the model's previous response rather than appending a new feedback message. The process consists of two stages:

1. In-Place Edit: The expert identifies an erroneous span within the model's response, directly modifies it, and prunes the downstream portion of the text that depends on the corrected span.
2. Continuation Generation: The LLM regenerates only the remainder of the response, continuing from the corrected context.

Unlike standard multi-turn feedback, which appends instructions to the dialogue history, or memory-based approaches that summarize past turns, in-place feedback removes the erroneous span from the conditioning context entirely. This prevents incorrect context from accumulating and keeps the effective input shorter.

The study validates this method across three complementary dimensions:

• Automated Benchmarks: Evaluation on five reasoning-intensive datasets (MATH-hard, MMLU-pro, GPQA, LiveCodeBench, and ZebraLogic) using LLM-based agents to simulate expert feedback.
• Fine-Grained Analysis: A turn-level decomposition on ZebraLogic puzzles to measure specific failure modes (retention of correct content, acceptance of feedback, and autonomous correction).
• User Study: A controlled study with 15 domain experts refining scientific summaries to measure satisfaction, fatigue, and perceived reliability.

Key Contributions and Results

1. Performance and Token Efficiency

Across all five benchmarks and five tested LLMs (ranging from 4B to 70B parameters), in-place feedback consistently outperforms standard multi-turn and memory-based feedback.

• Accuracy: In-place feedback achieves higher accuracy and exhibits faster improvement over turns. For example, on GPQA with Gemma, in-place feedback improved accuracy by 47.6 percentage points over the initial response after 10 turns, more than double the improvement of multi-turn feedback.
• Token Efficiency: Multi-turn feedback causes input tokens to grow linearly with each turn as the full history is appended. In-place feedback maintains a stable input length and generates significantly fewer tokens because it only regenerates the necessary continuation rather than the full response.

2. Mechanism of Improvement (Failure Mode Analysis)

Using the ZebraLogic dataset, the authors introduced three metrics to diagnose failure modes:

• Correctness-Preserving Ratio (CPR): Retention of previously correct content.
• Feedback Acceptance Ratio (FAR): Uptake of explicit corrective signals.
• Correction Through Reasoning Ratio (CTRR): Propagation of corrections to unmentioned parts of the reasoning.

Findings:

• Multi-turn Feedback: While CPR remains high (models retain correct content), FAR and CTRR decline steadily as turns accumulate. The accumulated history helps models reproduce prior content but stifles their ability to reason beyond explicit feedback.
• In-Place Feedback: Achieves significantly higher FAR and CTRR. By regenerating directly from the corrected span, the model avoids interference from stale reasoning and robustly propagates corrections to subsequent steps.

3. User Study Findings

In a study involving 15 experts refining scientific paper summaries:

• Satisfaction and Fatigue: Participants reported significantly higher final-output satisfaction and substantially lower fatigue with in-place feedback compared to multi-turn feedback.
• Qualitative Feedback: Experts noted that multi-turn feedback often led to "untargeted parts get[ting] modified" and "previously accepted edits [being] undone," requiring re-verification of the entire draft. In-place feedback bounded the re-verification region to the edited span.
• Mixed Strategy: A hybrid workflow, where participants chose between multi-turn and in-place feedback based on the task, received the highest ratings across all dimensions. Experts described using multi-turn for restructuring and in-place for polishing details.

Significance and Claims

The paper claims that editing errors directly is a more effective paradigm for expert-LLM collaboration than appending feedback instructions. The significance of in-place feedback lies in its ability to:

1. Mitigate Context Pollution: By removing erroneous spans from the context, it prevents the accumulation of incorrect information that confuses the model.
2. Enhance Reliability: It ensures that corrections are applied more reliably and propagated to subsequent reasoning steps without corrupting valid content.
3. Reduce Expert Burden: It lowers the cognitive load and fatigue associated with verifying full regenerations, making the refinement cycle more efficient.

The authors recommend in-place feedback as a sensible default for expert-LLM collaboration, while suggesting that exposing standard multi-turn feedback alongside it allows experts to select the most effective tool for specific correction types.

Limitations Noted by the Authors:

• The empirical evaluation focuses on models feasible within an academic computing budget; closed-source or significantly larger models were not tested.
• The automated evaluation relies on an LLM-based "in-place agent" to simulate edits, which is imperfect (producing appropriate edits only ~40-54% of the time). The authors note that real human edits are likely cleaner, suggesting the reported gains might be conservative.
• The user study was limited to a scientific summarization task with 15 participants; broader claims would require studies across more diverse domains and longer timeframes.
• The method currently depends on text-continuation APIs, which are not universally available from commercial providers.