TL;DR
Researchers have proposed the Polarity Composition Hypothesis, a high-level causal model explaining how Large Language Models (LLMs) process negation in Natural Language Inference (NLI). By using causal intervention techniques, they identified specific neural subspaces that represent event matching and polarity, proving that models like Qwen3 and Phi-4 solve these tasks by modularly separating "what happened" from "whether it is true or false."
What Problem Does This Paper Address?
Natural Language Inference (NLI) requires a model to determine if a hypothesis follows from, contradicts, or is unrelated to a premise. While LLMs have become quite good at this, they are notoriously brittle when it comes to negation. A single "not" can flip a label from Entailment to Contradiction, yet models often rely on superficial lexical shortcuts rather than actual logical reasoning.
The fundamental challenge is understanding the "black box": does a model have an internal mechanism for handling negation, or is it just memorizing patterns? This paper seeks to move beyond behavioral observations (looking at outputs) to causal evidence (looking at internal representations).
Methods
The authors propose a two-step causal model called the Polarity Composition Hypothesis:
- Event Matching ($m_i$): The model first determines if the content of an assumption (e.g., "John visited PlaceA") matches the claim.
- Relative Polarity ($\rho_i$): If the events match, the model compares their polarities (affirmative vs. negative). If they agree, the answer is TRUE; if they disagree, it's FALSE. If the events don't match, the polarity shouldn't matter, and the answer is UNKNOWN.
To test this, they used Distributed Alignment Search (DAS). DAS is a technique that identifies specific low-rank subspaces within a model's hidden states that correspond to high-level variables (like "polarity"). They then performed interchange interventions: swapping the "polarity" representation from one sentence into the hidden state of another to see if the model's final answer changed exactly as the hypothesis predicted.
Datasets
- Controlled SVO Dataset: A newly created synthetic dataset using a simple Subject-Verb-Object grammar (e.g., "John did not visit PlaceA"). This allowed for precise control over event matching and negation.
- MNLI (Multi-Genre NLI): A standard, broad-coverage NLI benchmark used to test if the causal mechanisms discovered in synthetic data transfer to natural, "in-the-wild" language.
Key Insights and Ideas
- Modular Processing: LLMs appear to separate the content of an event from its polarity. The researchers found distinct neural subspaces for each, suggesting the models aren't just processing sentences as holistic blobs of text.
- The "Gate" Mechanism: Event matching acts as a causal gate. Polarity information only flows through to the final decision if the event matching variable ($m_i$) is "open." If the events don't match, the model effectively ignores the polarity subspace.
- Abstraction over Surface Form: The identified "relative polarity" subspace is surprisingly robust. A subspace trained to recognize negation via "did not" successfully transferred to other forms like "never," "failed to," or "it is not the case that."
- Layer-wise Localization: Raw polarity (is this specific sentence negative?) is identified early in the model, near where the tokens are processed. However, the more abstract "relative polarity" (do these two sentences disagree?) persists into the much later layers and the final "answer" token.
Key Results
- High Causal Alignment: In Qwen3-8B and Phi-4 Mini, the researchers identified subspaces for relative polarity that achieved nearly 99% Interchange Intervention Accuracy (IIA). This means that by only touching a tiny fraction of the model's neurons, they could reliably flip the model's logical conclusion.
- Behavioral Consistency: Models showed a clear preference for the "Unknown" label when events mismatched, with a margin ($G_U$) consistently above 0, confirming the "gate" hypothesis behaviorally before the causal tests even began.
- MNLI Transfer: The subspaces learned on simple synthetic data partially transferred to the complex MNLI dataset, with T/F directionality preserved in 84-94% of cases for same-polarity interventions.
Limitations and Future Work
- Grammatical Simplicity: The study focused on simple SVO sentences. It does not yet account for more complex linguistic phenomena like quantifiers ("No one visited...") or nested negations.
- Single Match Assumption: The current hypothesis assumes at most one assumption event matches the claim. Real-world reasoning often involves aggregating multiple conflicting or supporting premises.
- Model Scope: The study was limited to two specific open-weight models (Qwen3 and Phi-4). While the results are compelling, it remains to be seen if larger models or different architectures (like encoders) utilize the same internal logic.
The authors suggest that future work should explore how these "polarity gates" interact with more complex semantic relations like lexical entailment (e.g., "walking" entailing "moving").
Summary updated