InternVL-X: Advancing and Accelerating the InternVL Series with Efficient Visual Token Compression

Visual tokens treated as a text prefix make LLM prefill and KV memory scale with image area, dominating VLM cost.

Three stacked compressions: (1) PVTC projector - integrates adjacent visual embeddings into local queries and uses a transformed CLS token as a global query, then point-to-region cross-attention; (2) LVTC - layer-wise token compression that shrinks the sequence in the LLM's SHALLOW layers and expands it back by upsampling plus residual in deeper layers, so cost falls without losing the depth needed for reasoning; (3) RVTC - dynamic high-resolution slicing that adjusts token count by image area.

Using 20% OR FEWER visual tokens, reports SOTA results on 7 public MLLM benchmarks and improves the average metric by 2.34% across 12 tasks versus InternVL. Baseline is the InternVL model and the token budget is stated - this is one of the few vision-token papers with an explicit baseline and budget.

Technical report, no peer-reviewed venue found. The LVTC idea - compress early, expand late with a residual - is the most interesting mechanism in this cluster for serving because it acknowledges that the deep layers are where vision tokens matter for reasoning while the shallow layers are where they are most redundant. Compressing only in shallow layers also shortens the KV cache for the whole remaining stack, which is the same effect as prefix-cache-friendly short prompts. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'unknown'. `venue` is 'arxiv-preprint', which means no peer-reviewed venue has been established for this work as of the verification date, so the track genuinely is not known. 'unknown' is used here rather than leaving the field null or guessing 'main': the schema separates 'unknown' from 'main' precisely so that an unprereviewed preprint is not silently counted as main-track work. This annotation is what lets a downstream filter distinguish 'peer reviewed, main track' from 'not peer reviewed' without inferring it from the venue string.