Conflict and Congruency Effects in Large Language Models: In-Weight and In-Context Competition in a Verbal Conflict Task

Direct Answer

The paper introduces a purely verbal conflict task for large language models (LLMs) that isolates how default word‑association tendencies compete with rule‑based instructions, revealing robust congruency effects similar to classic Stroop and flanker experiments. This matters because it provides a transparent, controllable probe of the internal “default vs. rule‑governed” dynamics that shape LLM behavior in real‑world prompting scenarios.

Background: Why This Problem Is Hard

Human cognition has long been studied with conflict tasks—Stroop, flanker, and Simon paradigms—that expose the tug‑of‑war between automatic responses and task‑relevant rules. Translating that insight to artificial neural networks is non‑trivial for three reasons.

• Opaque internal representations. LLMs store billions of parameters in a distributed fashion, making it difficult to pinpoint where a “default” mapping lives versus a “rule‑based” mapping.
• Prompt‑level variability. Existing benchmarks typically evaluate LLMs on static question‑answer pairs, which blend the influence of training data with the immediate prompt, obscuring the source of any observed conflict.
• Lack of mechanistic probes. Most prior work measures performance differences but does not trace the flow of attention or gradient signals that could explain why a model chooses one response over another.

Because LLMs are increasingly deployed as autonomous agents—customer‑support bots, code assistants, and decision‑support tools—understanding how they resolve competing cues is essential for safety, reliability, and interpretability.

What the Researchers Propose

The authors design a verbal conflict task that mirrors classic psychological experiments while staying entirely within the language domain. The task consists of three parts:

1. A prompt stem that implicitly encourages a “same‑color” completion (e.g., “The sky is ___”).
2. An explicit rule prefix that either aligns with the stem (congruent) or contradicts it (incongruent), such as “Answer with the opposite color:”.
3. A completion slot where the model must generate a color word.

Key components of the framework include:

• Default mapping. The statistical bias learned during pre‑training that links certain contexts (like “sky”) to a prototypical color (“blue”).
• Rule‑based mapping. The in‑context instruction that tells the model to override the default and follow a new mapping (e.g., “use the opposite color”).
• Attention pathways. The study hypothesizes two distinct routes: short‑range attention that latches onto the immediate color cue, and long‑range attention that reaches back to the rule prefix.

How It Works in Practice

When a model receives the prompt, its transformer layers compute attention scores across all tokens. The researchers track these scores to see which tokens dominate the decision process.

Conceptual Workflow

1. Tokenization. The prompt is split into sub‑word tokens, preserving the rule prefix, stem, and any filler words.
2. Initial attention sweep. Early layers tend to allocate high weight to the most recent tokens (the color cue), forming the short‑range pathway.
3. Contextual integration. Deeper layers broaden the receptive field, allowing attention to jump to the rule prefix, establishing the long‑range pathway.
4. Logit formation. The model aggregates contributions from both pathways; the dominant pathway determines the final token probability distribution.
5. Generation. A greedy or sampling decoder selects the color word, completing the task.

What sets this approach apart is the systematic manipulation of the rule prefix length and content, combined with causal attribution (e.g., gradient‑based token importance) and targeted attention ablations. By disabling short‑range or long‑range attention, the authors can directly observe performance shifts, confirming the existence of two competing routes.

Evaluation & Results

The authors evaluated seven LLM families:

• Gemma‑2‑2B
• Pythia models ranging from 410 M to 12 B parameters

Each model was tested on three conditions:

1. Baseline. No explicit rule; the model simply completes the stem.
2. Congruent. The rule prefix reinforces the default mapping.
3. Incongruent. The rule prefix demands the opposite color.

Key Findings

• Strong default same‑color tendency. Across all models, the baseline completion matched the most common color associated with the stem (e.g., “sky → blue”).
• Robust congruency effects. Six of the seven models performed significantly better on congruent trials than on incongruent ones, mirroring human Stroop interference.
• Distinct attention pathways. Causal attribution and attention heatmaps revealed that congruent trials relied heavily on short‑range attention to the immediate cue, while incongruent trials activated long‑range attention to the rule prefix.
• Fine‑tuning impact. Strengthening the default mapping via additional same‑color data improved congruent performance but degraded incongruent performance, confirming a trade‑off between the two pathways.
• Rule‑set size effect. Expanding the rule vocabulary (adding more varied instructions) selectively harmed incongruent trials, suggesting that larger rule sets dilute the long‑range pathway’s influence.

These results collectively support the authors’ central claim: congruency effects in LLMs arise from competition between an in‑weight default mapping and an in‑context rule‑based mapping, each mediated by separate attention mechanisms.

Why This Matters for AI Systems and Agents

Understanding the dual‑pathway dynamics has immediate practical implications for anyone building AI‑driven agents.

• Prompt engineering. Knowing that short‑range attention dominates unless a rule is sufficiently salient helps designers craft clearer instructions, reducing unintended defaults.
• Safety and alignment. If a model’s default mapping conflicts with a safety rule, the long‑range pathway may be too weak to override it, leading to hazardous outputs. Engineers can mitigate this by reinforcing rule‑based pathways during fine‑tuning.
• Agent orchestration. Multi‑agent systems often rely on hierarchical prompts (high‑level goals → low‑level actions). The study suggests that hierarchical prompting must respect the attention bandwidth of each layer to ensure higher‑level rules are not drowned out.
• Evaluation benchmarks. Traditional accuracy metrics miss the nuance of conflict resolution. Incorporating verbal conflict tasks into evaluation suites provides a richer picture of an agent’s controllability.

For teams using the UBOS platform overview to deploy LLM‑powered workflows, these insights can guide the selection of model sizes, fine‑tuning strategies, and prompt templates that balance default knowledge with task‑specific constraints.

What Comes Next

While the study opens a promising avenue, several limitations remain.

• Scope of language. The task is limited to color words; extending to richer semantic domains (e.g., sentiment, factual recall) will test the generality of the two‑pathway model.
• Model diversity. Only decoder‑only transformers were examined. Encoder‑decoder and multimodal models may exhibit different competition dynamics.
• Real‑time adaptation. The experiments use static prompts. Future work could explore how agents adapt their pathways during prolonged interactions or reinforcement‑learning loops.

Potential research directions include:

1. Designing adaptive prompting systems that monitor attention patterns in real time and inject clarifying rules when short‑range dominance is detected.
2. Developing fine‑tuning curricula that explicitly balance default and rule‑based pathways, perhaps using contrastive loss functions that penalize over‑reliance on one route.
3. Integrating the verbal conflict task into Workflow automation studio as a diagnostic step for new LLM deployments.

Ultimately, treating LLMs as experimental model systems for cognitive conflict could bridge AI research with decades of psychological theory, yielding agents that are both more predictable and more controllable.

References

Hu, X., Angstadt, M., Storks, S., Huang, Z., Taxali, A., Weigard, A., Lewis, R. L., & Sripada, C. (2026). Conflict and Congruency Effects in Large Language Models: In-Weight and In-Context Competition in a Verbal Conflict Task.