Modeling Multi-Dimensional Cognitive States in Large Language Models under Cognitive Crowding¶

Conference: ACL 2026 arXiv: 2604.17174 Code: GitHub Area: LLM Evaluation Keywords: cognitive state modeling, cognitive crowding, hyperbolic space, multi-dimensional joint prediction, CognitiveBench

TL;DR¶

This paper identifies that LLMs suffer a dramatic accuracy drop to 5.7% when jointly predicting four cognitive dimensions—sentiment, thinking style, stance, and intent—a phenomenon termed "cognitive crowding." Through Gromov \(\delta\)-hyperbolicity analysis, the paper demonstrates that cognitive states exhibit hierarchical structure, and proposes HyCoLLM, a framework that models cognitive states in hyperbolic space. An 8B model trained under this framework surpasses GPT-4o.

Background & Motivation¶

Background: LLMs perform well on individual tasks such as sentiment analysis, stance detection, and intent recognition, but these tasks are typically handled in isolation. Psychological research indicates that cognitive dimensions form an interactive system—for instance, an opposing stance may originate from deliberate analytical reasoning or from negative emotional states.

Limitations of Prior Work: (1) Existing benchmarks cover at most two cognitive dimensions (e.g., stance + sentiment), precluding the study of four-dimensional interactions. (2) The "thinking style" dimension—a critical bridge linking emotion to stance—lacks annotation in prior work. (3) LLMs perform well on single tasks but suffer severe performance degradation in joint multi-dimensional modeling; GPT-4o achieves only 5.7% joint accuracy across four dimensions.

Key Challenge: Cognitive states exhibit hierarchical/tree-like structure (Gromov \(\delta \approx 1\%\)), requiring exponentially growing representational space, whereas the Euclidean representation space of LLMs grows only polynomially. This "cognitive crowding" causes distinct cognitive states to overlap and become indistinguishable in Euclidean space.

Goal: (1) Construct CognitiveBench, the first four-dimensional cognitive benchmark. (2) Diagnose and explain the joint modeling bottleneck in LLMs. (3) Propose a geometry-aware solution.

Key Insight: Leverage the natural exponential volume growth and hierarchical support of hyperbolic space to alleviate cognitive crowding.

Core Idea: Model cognitive states in hyperbolic space (Poincaré ball), separate distinct states via geometry-aware contrastive loss, and align LLMs' internal representations to this structure through Hyperbolic Guided Alignment Tuning.

Method¶

Overall Architecture¶

HyCoLLM operates in two stages: (1) Hyperbolic Cognitive Network (HCN)—learns cognitive state embeddings on the Poincaré ball using a geometry-aware contrastive loss to separate distinct states; (2) Hyperbolic Guided Alignment Tuning (HGAT)—aligns the LLM's internal representations to the learned hyperbolic cognitive manifold via a semantic-cognitive topology loss.

Key Designs¶

CognitiveBench Dataset:
- Function: The first benchmark annotated across four cognitive dimensions (sentiment, thinking style, stance, and intent).
- Mechanism: Posts on four topics (China–US trade, U.S. elections, DEI, and Federal Reserve interest rates) were collected from Twitter and filtered through a multi-stage pipeline to yield approximately 9,000 candidate samples. Twenty-nine annotators with backgrounds in psychology and affective computing independently labeled each sample (three annotators per sample); only samples with at least two-thirds agreement were retained, yielding a final dataset of 6,514 entries.
- Design Motivation: The four-dimensional label schema is grounded in established psychological theories—Plutchik's emotion model, dual-process theory (intuitive vs. analytical thinking), social judgment theory (stance), and speech act theory (intent).
Hyperbolic Cognitive Network (HCN):
- Function: Learns well-separated cognitive state embeddings in hyperbolic space.
- Mechanism: Sentence embeddings are mapped onto the Poincaré ball; a geometry-aware contrastive loss pulls together embeddings of the same cognitive state and pushes apart those of different states. The exponential volume growth of hyperbolic space provides sufficient capacity to accommodate all state combinations (9×8×3×7 = 1,512 combinations).
- Design Motivation: Euclidean space cannot effectively separate 1,512 cognitive states due to its polynomial volume growth, whereas hyperbolic space, whose volume grows exponentially with radius, is naturally suited to hierarchically structured data.
Hyperbolic Guided Alignment Tuning (HGAT):
- Function: Aligns the LLM's internal representations to the hyperbolic cognitive manifold.
- Mechanism: A Semantic-Cognitive Topology Loss is designed to constrain the topological structure of the LLM's hidden states during fine-tuning to be consistent with the hyperbolic cognitive space learned by HCN, enabling the LLM to exploit hierarchical relationships among cognitive states during inference.
- Design Motivation: Learning cognitive embeddings in hyperbolic space alone is insufficient; the geometric prior must be injected into the LLM's reasoning process.

Loss & Training¶

HCN employs a hyperbolic contrastive loss. HGAT uses a Semantic-Cognitive Topology Loss combined with a standard generation loss. The backbone model is LLaMA-3.1-8B-Instruct.

Key Experimental Results¶

Main Results¶

Model	Avg. Single-Dimension Accuracy	Four-Dimension Joint Accuracy
GPT-4o	~50–60%	5.7%
LLaMA-3.1-8B (SFT)	~45–55%	~4%
HyCoLLM-8B	Improved	Significantly improved (surpasses GPT-4o)

Ablation Study¶

Configuration	Joint Accuracy	Notes
HyCoLLM (Full)	Highest	Complete framework
w/o HCN	Degraded	Without hyperbolic cognitive network
w/o HGAT	Degraded	Without alignment tuning
Euclidean contrastive learning	Degraded	Validates necessity of hyperbolic geometry

Key Findings¶

GPT-4o achieves reasonable single-dimension performance but only 5.7% on the four-dimensional joint task—this reflects a geometric limitation of the representation space rather than a lack of capability.
Gromov \(\delta\)-hyperbolicity analysis confirms that CognitiveBench exhibits a relative \(\delta \approx 1\%\), indicating strong hierarchical structure.
HyCoLLM's 8B model surpasses GPT-4o on joint modeling, demonstrating the effectiveness of geometric priors.
Incorporating the thinking style dimension substantially affects the prediction of stance and intent, confirming genuine interactions among the four cognitive dimensions.

Highlights & Insights¶

The concept of "cognitive crowding" precisely diagnoses the bottleneck of multi-dimensional joint modeling in LLMs—framing it as a geometric limitation rather than a capability deficiency.
Applying Gromov \(\delta\)-hyperbolicity analysis to characterize data structure provides a data-driven criterion for determining when hyperbolic space is appropriate.
The result that an 8B model surpasses GPT-4o powerfully demonstrates the value of geometric priors.

Limitations & Future Work¶

CognitiveBench covers only English Twitter data; cross-cultural and cross-lingual generalization remains unknown.
Hyperbolic space operations increase training complexity and risk of numerical instability.
The four-dimensional label schema may still be incomplete—deeper cognitive dimensions such as personality and values are not addressed.
The annotation process is costly (29 experts over two months), limiting scalability.

vs. SemEval-16: Covers only stance and sentiment, without the thinking style dimension.
vs. DoT (Chen et al.): DoT focuses on detecting a single type of cognitive distortion, whereas this paper performs joint modeling across multiple dimensions.
vs. Hyperbolic Embeddings: Prior applications of hyperbolic space in NLP have been primarily limited to word embeddings and knowledge graphs; this paper represents the first application to cognitive state modeling.

Rating¶

Novelty: ⭐⭐⭐⭐⭐ — Highly original: cognitive crowding concept + hyperbolic space solution + four-dimensional benchmark.
Experimental Thoroughness: ⭐⭐⭐⭐ — Ablations are comprehensive, though only a single backbone model is evaluated.
Writing Quality: ⭐⭐⭐⭐ — Framework is clearly presented, though some technical sections are dense.
Value: ⭐⭐⭐⭐⭐ — Reveals a fundamental bottleneck in LLM cognitive modeling and provides an effective solution.