Abstract
This research, presented at EMNLP 2023, investigates whether Large Language Models (LLMs) possess "analytic intelligence" comparable to human abstract reasoning. By introducing Blackbird Language Matrices (BLM)—a linguistic adaptation of Raven’s Progressive Matrices—the study evaluates the ability of models like GPT-4 to perform out-of-distribution generalization. The findings reveal a significant gap: while LLMs are proficient at identifying linguistic components (correspondence finding), they frequently fail to induce the global rules governing a sequence, often defaulting to surface-level heuristics.
Introduction: Beyond Statistical Pattern Matching
In machine learning, generalization typically refers to a model's performance on new data drawn from the same distribution as its training set. However, human intelligence is characterized by the ability to extract abstract rules from a few examples and apply them to entirely novel contexts. To test this in LLMs, the author adapts a classic tool from cognitive psychology: Raven’s Progressive Matrices (RPMs).
In a standard RPM, a subject must identify rules (e.g., rotation, color change) within a 3x3 grid of shapes. The BLM task translates these visual puzzles into linguistic ones to determine if LLMs can follow abstract grammatical rules rather than just matching statistical patterns.
The Blackbird Language Matrices (BLM) Task
The BLM benchmark consists of matrices where sentences follow specific linguistic rules. A model is presented with a "context" of seven sentences and must select the eighth sentence from a set of foils.
The benchmark focuses on several linguistic phenomena:
- Subject-Verb Agreement: Rules where verb number must match the subject despite intervening "attractor" phrases.
- Causative Alternation: Rules for verbs that can be both transitive and intransitive (e.g., "The boy opened the door" vs. "The door opened").
- Sequence of Tenses: Complex Italian grammar involving conditional and subjunctive moods.
- Mixture Matrices: High-difficulty problems requiring the model to track two rules simultaneously.
The data is further categorized into Type I (repetitive vocabulary) and Type III (varied vocabulary) to test if lexical diversity interferes with rule induction.
Methodology: Evaluating Analytic Intelligence
The study employed both qualitative and quantitative analyses. Qualitative "think-aloud" sessions were conducted with GPT-3.5, using prompts that required the model to explain its hypotheses step-by-step. Quantitatively, GPT-4 was tested on a controlled sample of 30+ problems using a coding protocol inspired by cognitive science to evaluate three specific skills:
- Correspondence Finding: Identifying relevant words (objects) and grammatical features (attributes).
- Goal and Subgoal Management: Organizing the search for a rule (e.g., checking one attribute at a time).
- Item Novelty: Handling vocabulary not previously seen in the current context.
Key Insights: The Disconnect Between Identification and Induction
The results highlight a fascinating disconnect in LLM reasoning. GPT-4 demonstrated high proficiency in "Correspondence Finding," correctly identifying subjects, verbs, and their attributes with 73-89% accuracy. However, it struggled to synthesize these parts into a global rule.
In the causative alternation task, GPT-4 identified the correct underlying rule in only one out of eleven cases, even when it selected the correct answer. This suggests that LLMs often rely on "semantic consistency" heuristics. For instance, if context sentences involve "doctors," the model may pick a foil about "doctors" because it feels related, even if that sentence violates the matrix's grammatical rule.
Furthermore, when faced with "Mixture Matrices" involving two simultaneous rules, the models' reasoning often collapsed. They might follow one rule (e.g., tense) while completely ignoring another (e.g., agreement), indicating a limitation in managing multiple subgoals within "working memory."
Results and Limitations
While GPT-4's success rate in picking the correct answer was high (e.g., 11/11 for Causatives), its rate of correctly identifying the actual pattern was significantly lower. Additionally, increasing lexical variety (Type III) made the tasks much harder for the models, whereas humans are generally robust to such changes once a rule is understood.
The study's limitations include a relatively small quantitative sample size (33 problems) and a focus primarily on the GPT family. Future work should expand BLM to more languages and model architectures, and potentially explore internal model states to see if these rules are represented mechanistically within the neural network.
Conclusion
The BLM benchmark reveals that while current LLMs are adept at linguistic decomposition, they lack the robust "analytic intelligence" required for true rule-based induction. Their reliance on surface-level heuristics over logical consistency suggests that achieving human-like abstract reasoning remains a significant challenge for neural network architectures.
Summary updated