TELLER: Dual-Path Iterative Preference Optimization for Table Entity Linking
TELLER introduces a dual-path framework for table entity linking that combines iterative direct preference optimization with iterative length-normalized regularized preference optimization for reasoning The direct-answer path refreshes preference data with residual errors from the updated model, achieving accuracy gains from 94.35% to 94.50% on TableInstruct and 87.59% to 88.20% on MammoTab V2 The reasoning path uses filtered and compressed chain-of-thought rationales, improving accuracy from 92
Analysis
TL;DR
- TELLER introduces a dual-path framework for table entity linking that combines iterative direct preference optimization with iterative length-normalized regularized preference optimization for reasoning
- The direct-answer path refreshes preference data with residual errors from the updated model, achieving accuracy gains from 94.35% to 94.50% on TableInstruct and 87.59% to 88.20% on MammoTab V2
- The reasoning path uses filtered and compressed chain-of-thought rationales, improving accuracy from 92.90% to 92.95% on TableInstruct and from 79.09% to 81.85% on MammoTab V2 while maintaining high complete reasoning generation rates
- The key innovation addresses static training supervision by making preference data adaptive to the evolving model's residual errors
- Both concise entity prediction and explicit reasoning benefit from iterative preference learning, demonstrating the framework's versatility
Why It Matters
This work addresses a critical limitation in current entity linking systems: static preference data that cannot adapt to the residual errors of evolving models. For AI practitioners building table-based information extraction pipelines, TELLER offers a practical dual-path approach that improves accuracy on both direct prediction and reasoning-intensive tasks. The iterative preference optimization framework could serve as a template for other structured data understanding applications beyond entity linking.
Technical Details
- Dual-path architecture: TELLER employs two parallel paths—a direct-answer path using iterative direct preference optimization (DPO) with residual-error-driven preference data refresh, and a reasoning path using filtered/compressed chain-of-thought rationales followed by iterative length-normalized regularized preference optimization
- Candidate retrieval: The system first retrieves and ranks Wikidata candidates, then retains reduced table evidence in the prompt to balance context richness with computational efficiency
- Iterative preference learning: Unlike static preference datasets, TELLER dynamically refreshes preference data based on the residual errors produced by the updated model at each iteration, enabling continuous adaptation
- Length normalization: The reasoning path incorporates length-normalized regularized preference optimization to address bias introduced by variations in reasoning length during sequence-level preference learning
- Benchmarks: Evaluated on TableInstruct entity-linking subset and MammoTab V2, with the direct-answer path showing +0.15% and +0.61% accuracy improvements, and the reasoning path showing +0.05% and +2.76% improvements respectively
Industry Insight
- The iterative preference optimization approach demonstrated in TELLER could be generalized to other structured data tasks where static training data fails to capture model evolution, suggesting a broader paradigm shift from one-shot preference learning to continuous adaptive training
- The dual-path design—separating concise prediction from explicit reasoning—offers a practical blueprint for systems that need to balance latency (direct answers) with interpretability (reasoning traces), particularly in domain-specific applications like healthcare or finance where explainability matters
- The significant reasoning path improvement on MammoTab V2 (+2.76%) versus the smaller gain on TableInstruct suggests that iterative preference learning with reasoning is particularly valuable for complex, domain-specific table linking tasks, warranting targeted investment in reasoning-capable pipelines for challenging real-world datasets
Disclaimer: The above content is generated by AI and is for reference only.