Hongxu Zhou

Hongxu Zhou

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.

Hongxu Zhou

State space models (SSMs) have been shown to possess the theoretical capacity to model both star-free sequential tasks and bounded hierarchical structures Sarrof et al. (2024). However, formal expressivity results do not guarantee that gradient-based optimisation will reliably discover the corresponding solutions. Existing benchmarks probe either monotonic state tracking, as in the standard Flip-Flop task, or structural nesting, as in the Dyck languages, but neither isolates reversible semantic state retrieval. We introduce the UNDO Flip-Flop task to fill this gap. By extending the standard Flip-Flop with an UNDO, the task requires a model to maintain an implicit bounded stack and recover historical states under non-monotonic update sequences. We evaluate one-layer and two-layer Mamba-2 under this framework. Both variants fail to acquire the provably expressible stack-based rollback mechanism, converging instead on a local toggle heuristic that inverts the current state rather than retrieving stored history. Under an adversarial retraction pressure test held within the training length distribution, the two-layer model collapses to 41.10% accuracy, which is below random chance. The results confirm systematic rather than incidental failure. Causal ablation shows that the bottleneck lies in retrieval, not storage. These results draw a clear line between what an architecture can in principle represent and what gradient descent reliably learns, a distinction that theoretical expressivity analyses alone cannot capture.

Hongxu Zhou, Hylke Westerdijk, Khondoker Ittehadul Islam

This study investigates how context and emotional tone metadata influence large language model (LLM) reasoning and performance in fallacy classification tasks, particularly within political debate settings. Using data from U.S. presidential debates, we classify six fallacy types through various prompting strategies applied to the Qwen-3 (8B) model. We introduce two theoretically grounded Chain-of-Thought frameworks: Pragma-Dialectics and the Periodic Table of Arguments, and evaluate their effectiveness against a baseline prompt under three input settings: text-only, text with context, and text with both context and audio-based emotional tone metadata. Results suggest that while theoretical prompting can improve interpretability and, in some cases, accuracy, the addition of context and especially emotional tone metadata often leads to lowered performance. Emotional tone metadata biases the model toward labeling statements as \textit{Appeal to Emotion}, worsening logical reasoning. Overall, basic prompts often outperformed enhanced ones, suggesting that attention dilution from added inputs may worsen rather than improve fallacy classification in LLMs.