From Hallucination to Structure Snowballing: The Alignment Tax of Constrained Decoding in LLM Reflection

TL;DR

This paper examines how constrained decoding in large language models can inadvertently cause a new failure mode called 'structure snowballing,' highlighting an inherent trade-off between structural control and model capacity.

Contribution

It demonstrates that structural constraints alone do not improve self-correction and can lead to new errors, revealing an 'alignment tax' in constrained decoding methods.

Findings

Imposing structural constraints does not enhance self-correction performance.

Strict formatting rules increase cognitive load, causing formatting traps.

Structural constraints can trigger 'structure snowballing' failures.

Abstract

Intrinsic self-correction in Large Language Models (LLMs) frequently fails in open-ended reasoning tasks due to ``hallucination snowballing,'' a phenomenon in which models recursively justify early errors during free-text reflection. While structured feedback can mitigate this issue, existing approaches often rely on externally trained critics or symbolic tools, reducing agent autonomy. This study investigates whether enforcing structured reflection purely through Outlines-based constrained decoding can disrupt error propagation without additional training. Evaluating an 8-billion-parameter model (Qwen3-8B), we show that simply imposing structural constraints does not improve self-correction performance. Instead, it triggers a new failure mode termed ``structure snowballing.'' We find that the cognitive load required to satisfy strict formatting rules pushes the model into formatting traps. This observation helps explain why the agent achieves near-perfect superficial syntactic alignment yet fails to detect or resolve deeper semantic errors. These findings expose an ``alignment tax'' inherent to constrained decoding, highlighting a tension between structural granularity and internal model capacity in autonomous workflows. Code and raw logs are available in the GitHub repository: https://github.com/hongxuzhou/agentic_llm_structured_self_critique.