Abstract
Modern Natural Language Understanding (NLU) models often function as "black boxes," making it difficult to verify if they maintain a consistent internal world model during narrative processing. We present Breakpoint Transformers, a framework that enables models to track and report internal "beliefs" at arbitrary points in a text. By incorporating specialized breakpoint tokens and a single-read architecture, our approach achieves state-of-the-art results on complex reasoning tasks while significantly improving interpretability and computational efficiency.
Introduction
While Large Language Models (LLMs) excel at end-of-text question answering, their internal logic remains opaque. In narrative understanding—such as tracking an object's location across multiple sentences—it is unclear whether a model maintains a coherent state or relies on statistical shortcuts. Existing probing methods are often post-hoc and computationally expensive, requiring multiple passes over the input text.
To address these challenges, we introduce Breakpoint Modeling, inspired by the use of breakpoints in software debugging to inspect variable states at specific execution steps.
The Breakpoint Framework
The core of our approach is the insertion of special breakpoint tokens, [B], at intermediate points within a story (e.g., after every sentence). These tokens are trained to capture the "state of the world" at that specific moment.
Single-Read Architecture and Efficiency
A key innovation is the Single-Read design. Traditional "multi-pass" models must re-process the entire story for every belief query. In contrast, our model encodes the story once, extracting embeddings for all [B] tokens simultaneously.
To verify a belief (e.g., "Is the apple in the kitchen?"), the model uses Late Interaction: it encodes the query separately and applies a simple bilinear scorer against the relevant breakpoint embedding. This architecture reduced training time by 54% (from 63 to 34 hours) in our experiments without compromising accuracy.
Multi-task Learning
The model is optimized using three simultaneous objectives:
- Proposition Prediction: Determining if a statement is True, False, or Unknown at a given breakpoint.
- End-Task QA: Answering the final question (e.g., "Where is the apple?").
- Auxiliary Denoising: Reconstructing local context from breakpoint embeddings to ensure they contain rich, localized information.
Experimental Results
We evaluated the Breakpoint Transformer across three reasoning benchmarks: CLUTRR (relational reasoning), bAbI (entity tracking), and TRIP (intuitive physics).
Performance and Consistency
On the TRIP dataset, which requires identifying physical impossibilities, the Breakpoint Transformer achieved state-of-the-art results. Notably, it showed a 23-32% absolute improvement in consistency and verifiability over RoBERTa-based baselines. In CLUTRR, we observed a 4-8% boost in proposition accuracy. Crucially, adding intermediate belief tracking did not degrade performance on the primary QA tasks, suggesting that interpretability and accuracy are not mutually exclusive.
Logic and Future-Masking
Training with intermediate supervision encourages more logical internal states, resulting in fewer "global constraint violations" (e.g., the model no longer believes a person is in two locations at once). We also found that applying future-masking to the breakpoint self-attention layer prevents the model from "cheating" by looking ahead to the end of the story, ensuring the breakpoint accurately represents the state at that moment.
Conclusion and Future Work
Breakpoint Transformers offer a scalable path toward more interpretable and efficient NLU. While they significantly improve consistency, challenges remain. Models still struggle with compositional generalization to longer or more complex stories than those seen during training. Furthermore, while the model reports its beliefs, a "causal gap" remains: a correct belief at a breakpoint does not strictly guarantee that this belief caused the final correct answer. Future research will focus on bridging this gap and reducing the annotation cost associated with intermediate labels.
Summary updated