exp015 -- GTLM table-graph with BIDIRECTIONAL row chains
Public mirror of one experiment report from a private research repo.
Summary
Tests whether making cell-row chains BIDIRECTIONAL (every adjacent cell pair gets edges in BOTH directions) on the GTLM table-as-graph paradigm recovers the order-dependent question subtypes that collapsed in the prior left-only chain variant. Base: Llama-3.2-1B + LoRA. Train: TableInstruct DP subset. Eval: TableBench DP test, Claude Sonnet LLM-judge.
Verdict: null on the overall headline (-0.12pp paired n=834, McNemar p=1). Bidirectional rows partially recovered MatchBased (-15pp -> -4pp) but DescriptiveAnalysis stayed broken; CausalAnalysis (the prior win) was preserved at +8.9pp.
