The TIS 2.0 update introduces a passage reordering capability that eliminates the "lost-in-the-middle" position bias in retrieval-augmented generation without requiring additional training. By applying learned token importance scores to reorder retrieved documents, the system achieves zero position gap and improves exact match accuracy by 5 percentage points over baselines.

  • Reordering passages via TIS scores reduces the LITM gap to 0.000 and raises EM to 21.7%, matching oracle performance on MS-MARCO queries.
  • The existing `tis-stage3-ert` checkpoint, originally trained for KV cache compression, transfers zero-shot to passage reordering with identical results to a dedicated reranker.
  • A new `QueryAwareImportanceHead` module adds cross-attention query-context interaction and InfoNCE contrastive training for per-passage relevance scoring.
  • Integrating TIS importance bias into speculative decoding improves drafter acceptance length from 5.80 to 6.57, yielding a 12.5% speedup on LLaMA-3 models.
  • New checkpoints `tis-passage-reranker` and `tis-v8b-hard-anchor` are now available on HuggingFace alongside the updated GitHub repository.

The authors note that while TIS effectively addresses position bias, answer extraction remains the primary bottleneck, as approximately 76% of test cases remain incorrect even after optimal reordering.