Neurosymbolic Repair for Low-Code Formula Languages
summary
LaMirage generates repair engines from an annotated context free grammar plus small domain strategies, enumerates syntax-correct edits with a bounded backtracker, uses a lightweight pointer network to suggest distant error locations, and ranks candidates by edit distance plus a learned neural ranker. It targets last-mile fixes requiring few edits and evaluates on 400 real Excel and Power Fx formulas with manually annotated ground truth, outperforming state-of-the-art symbolic and neural baselines while guaranteeing grammatically valid repairs.
AI-written summary · about
chapters (12)
- 0:00Neurosymbolic repair intro
- 2:24Mirage engine overview
- 5:15Repair candidates enumeration
- 8:41Ranking repair candidates
- 10:05Evaluation on Excel & Power Fx
- 13:59Key insights
- 14:40Future work
- 15:02Q&A: Traditional error recovery
- 17:58Q&A: Low-code challenges
- 19:01Q&A: Copilot vs autocomplete
- 21:54Q&A: Parser choice
- 23:58Q&A: Extending to Python
related talks
01h 10mNeurosymbolic Program Synthesis: Bridging Perception and Reasoning in Real-World(…)
18mActive Learning for Neurosymbolic Program Synthesis
23mEnhancing Neural Network Robustness via Synthesis of Repair Programs
42mReasonable AI
12mLanguage-Model Probabilistic Programming for Improving Autoformalization via Cycle(…)
28mSymbolic Reasoning in telecom oriented LLMs
01h 00mBetter Learning through Programming Languages
55mLessons Learned From Implementing Generative AI Solutions