RTZeroShotSeqClassification (Zero-shot schema labeling + segmentation)

Idea

The RTZeroShotSeqClassification family segments a reasoning trace by first assigning a discourse / reasoning label to each base unit (sentence or clause) using zero-shot sequence classification, and then merging consecutive units that share the same predicted label.

This yields a segmentation that is explicitly interpretable: segment boundaries occur when the predicted schema role changes.

We provide:

• Generic zero-shot segmentation (RTZeroShotSeqClassification) with configurable label sets.
• Two specialized variants aligned with our paper’s schemas:
  • Reasoning Flow (RTZeroShotSeqClassificationRF)
  • Thought Anchor (RTZeroShotSeqClassificationTA)

---

Method (high-level)

Given a trace and a base unit choice (sent or clause):

1. Base segmentation Compute base offsets via:
   • SegBase.get_base_offsets(trace, seg_base_unit=...)
2. Zero-shot classification per base unit For each base span u_i, run a zero-shot NLI classifier:
   • label_i = argmax_label p(label | u_i)

   The implementation uses the HuggingFace pipeline("zero-shot-classification") with multi_label=False (forced single best label).

3. Merge adjacent spans with identical labels Consecutive base units are merged into a segment as long as their predicted labels remain the same. A new segment starts when label_i != label_{i-1}.

Output:

• final_offsets: merged character offsets
• final_labels: the predicted schema label for each merged segment

---

Models used

These engines use NLI-style zero-shot classification models via the HuggingFace pipeline.

Supported model names in the code:

• facebook/bart-large-mnli
• FacebookAI/roberta-large-mnli
• MoritzLaurer/mDeBERTa-v3-base-mnli-xnli

Implementation notes:

• Model is loaded with device_map="auto" and torch_dtype="auto".
• Inference is done independently per base unit (sentence/clause).

For reproducibility, report the exact model used and the exact label set used (below), since both directly determine the segmentation.

---

Label sets (explicit)

Generic default labels (RTZeroShotSeqClassification)

Context
Planning
Fact
Restatement
Example
Reflection
Conclusion

Reasoning Flow schema (RTZeroShotSeqClassificationRF)

Context
Planning
Fact
Reasoning
Restatement
Assumption
Example
Reflection
Conclusion

Thought Anchor schema (RTZeroShotSeqClassificationTA)

Problem Setup
Plan Generation
Fact Retrieval
Active Computation
Uncertainty Management
Result Consolidation
Self Checking
Final Answer Emission