ProCreations's picture
Publish validated ICML reproduction
5114c4e verified
Raw
History Blame Contribute Delete
1.21 kB
[
"Table data exhibits periodic non-vanishing mutual information at lags d = km (k = 1,2,...) for m-column tables, in contrast to the power-law decay of mutual information in natural language (Theorem 2.2).",
"Table mutual information asymptotically dominates natural-language mutual information at these periodic lags (Corollary 2.4), and the effective dependency distance is infinite for tables versus finite for text (Theorem 2.6).",
"TableLong's SQL-based pipeline, built from 10,000+ real-world tables via consistency-based filtration, improves average long-context benchmark performance by +8.24% for the DS-R1-Distill-32B model (Table 1).",
"On the out-of-domain LiveCodeBench benchmark, the 32B model improves by +11.97% after TableLong training, part of an average +8.06% out-of-domain gain (Table 4).",
"On Needle-in-a-Haystack retrieval, the 32B model improves from 87.95% to 99.40% (+13.02%) after training with TableLong data (Figure 3).",
"Ablations show multi-hop reasoning performance scales with table cell count (46.30% with ~30 cells to 48.36% with ~300+ cells, Table 2), and using multiple tables versus a single table improves grounding from 46.66% to 48.36% average (Table 3)."
]