| # Repro - Training-Inference Consistent Segmented Execution for Long-Context LLMs | |
| ## Pages | |
| | Page | | |
| | --- | | |
| | [Claim 1: Training and inference share identical segment-level forward semantics](#/claim-1-training-and-inference-share-identical-segment-level-forward-semantics) | | |
| | [Claim 2: TBPTT computes the exact gradient of the stated truncated objective](#/claim-2-tbptt-computes-the-exact-gradient-of-the-stated-truncated-objective) | | |
| | [Claim 3: The truncated training objective is aligned with segmented inference](#/claim-3-the-truncated-training-objective-is-aligned-with-segmented-inference) | | |
| | [Claim 4: Long-range retrieval is sparse and forward-only](#/claim-4-long-range-retrieval-is-sparse-and-forward-only) | | |
| | [Sources, protocol, and integrity](#/sources-protocol-and-integrity) | | |
| | [Conclusion](#/conclusion) | | |
| | [segmented-64223-repro](#/segmented-64223-repro) | | |