ICML 2026 · EMM-QA Workshop PMLR 306

Salient
Knowledge
Pathways

Sparse cross-modal routing for efficient knowledge-intensive multimodal question answering — matching or exceeding dense baselines with up to 6.8× fewer FLOPs and 2.7× less latency.

Noor Islam S. Mohammad*  ·  Uluğ Bayazıt
Istanbul Technical University, Türkiye
islam23@itu.edu.tr

4.07×
Fewer FLOPs than RA-CM3
2.69×
Faster end-to-end latency
+4.3
Accuracy pts on InfoSeek
5
KI-MMQA benchmarks, all SOTA
11%
Visual token retention at 99% no-pruning accuracy — matching O(1/√V) theory
0.42 T
FLOPs per query
(vs. 1.71T for RA-CM3)
305 ms
End-to-end latency
on A100-80GB
34–41%
Queries fast-pathed
via SKV drafter
9.8×
BSCA fusion
FLOP reduction
Venue
EMM-QA Workshop
ICML 2026, Seoul, South Korea
PMLR 306
Benchmarks
OK-VQA · A-OKVQA
InfoSeek · Encyclopedic-VQA
ViQuAE
Backbone
Vicuna-7B v1.5 (LoRA rank-64)
EVA-CLIP-G vision encoder
SPLADE++ retrieval
Code & Data
github.com/skip-mmqa
Checkpoints: CC-BY
FAISS index: 240 GB snapshot

Dense pricing paid unconditionally; sparse routing spent wisely.

Knowledge-intensive multimodal question answering (KI-MMQA) sits at the intersection of three expensive primitives: long visual token sequences, dense retrieval over large external corpora, and full cross-modal fusion. Existing systems pay all three costs uniformly per query, even though only a small fraction of visual content and retrieved knowledge is actually relevant to any given question.

We introduce SKIP (Salient Knowledge-Injected Pathways), a unified inference architecture that routes computation along sparse pathways jointly conditioned on the question, the image, and a difficulty estimate. SKIP combines question-guided visual token pruning, region-conditional sparse retrieval, bipartite sparse cross-attention, and speculative knowledge verification with an adaptive budget controller that allocates compute proportional to predicted question difficulty.

We derive an information-bottleneck bound showing that the optimal visual sparsity rate scales as O(1/√N) under realistic question-image mutual-information assumptions, with retained accuracy guarantees. Across five KI-MMQA benchmarks, SKIP matches or exceeds the accuracy of strong dense baselines while using 3.4–6.8× fewer FLOPs and 2.7× less end-to-end latency.

Five learned components, one sparse routing policy.

SKIP replaces the dense KI-MMQA pipeline with five components arranged so that each stage's sparsity decision conditions the next. The key commitment: sparsity is applied compositionally—each stage's output is the sparse input to the next, not independently sparsified.

Image I Question q VEnc EVA-CLIP QEnc T×d DBC budget (V′, K′) QVS prune V′ SKV fast path RCSR per-region Corpus 6.5M passages BSCA bipartite Decoder LLM Answer â fast path (34–41% of queries) data flow budget control routing/skip
COMPONENT 1
QVS
Question-Guided Visual Saliency
Cross-attention module scoring each visual token's relevance to the question. Prunes to top-V′ before retrieval, ensuring queries derive from question-relevant regions only.
sᵢ = σ(fq(Q)ᵀfv(vᵢ)/√d + wᵦᵀvᵢ)
COMPONENT 2
RCSR
Region-Conditional Sparse Retrieval
Clusters retained tokens into ≤8 salient regions via k-means and issues a separate SPLADE retrieval query per region, surfacing knowledge tied to small discriminative areas.
Recall@4: 43% → 71% on InfoSeek hard split
COMPONENT 3
BSCA
Bipartite Sparse Cross-Attention
Constructs a bipartite compatibility graph between visual tokens and knowledge chunks; computes attention only on edges exceeding threshold τ. Block-sparse FlashAttention kernel.
9.8× fewer fusion FLOPs · 7.2× layer speedup
COMPONENT 4
DBC
Difficulty-Adaptive Budget Controller
3-layer MLP predicting per-query compute budgets (V′, K′) from global image–question features. Concentrates capacity on hard queries, recovers it from easy ones.
Median: V′=64, K′=6 · avg difficulty 0.41±0.18
COMPONENT 5
SKV
Speculative Knowledge Verification
700M-parameter drafter that speculatively answers easy queries and short-circuits the full pipeline when confidence exceeds τ_SKV. Extends speculative decoding to knowledge access.
34–41% fast-pathed · ≤1% accuracy cost

SKIP Pareto-dominates all baselines at every FLOP budget.

OK-VQA Accuracy
All systems use 7B backbone · % accuracy
BLIP-2
45.9
LLaVA-1.5
50.3
KAT
54.4
RAVQA-2
56.8
ReVeaL
58.0
RA-CM3
60.5
SKIP
63.2
InfoSeek Accuracy
Entity-centric, hardest benchmark · exact match %
BLIP-2
11.2
LLaVA-1.5
13.7
KAT
18.6
RAVQA-2
21.3
ReVeaL
23.8
RA-CM3
26.4
SKIP
30.7
Accuracy (%) / TFLOPs / Latency (ms). All systems use 7B backbone. Bold = best.
Model OK-VQA A-OKVQA InfoSeek Enc-VQA ViQuAE TFLOPs Lat. (ms)
BLIP-2 no retrieval 45.939.711.28.419.1 0.42210
LLaVA-1.5 no retrieval 50.344.113.710.222.3 0.51240
KAT 54.448.218.614.728.9 0.93410
RAVQA-2 56.850.621.316.431.7 1.18520
ReVeaL 58.052.423.818.133.2 1.34610
RA-CM3 60.555.326.420.935.6 1.71820
SKIP ours 63.258.430.7 24.337.4 0.42305
vs. RA-CM3 +2.7+3.1+4.3+3.4+1.8 4.07× less2.69× faster

Each component removes a distinct category of distractor.

No individual technique recovers the full system effect. The gains compound: each sparse component filters a different source of noise, and their composition reaches 63.2% that no single component approaches.

OK-VQA · Component Ablation

Table 2
Full SKIP 63.2% 0.42T
− QVS (random pruning) 54.1% 0.42T
− QVS (no pruning) 62.4% 1.38T
− RCSR → global retrieval 59.7% 0.48T
− BSCA → dense 62.9% 1.21T
− DBC (fixed V′=64, K′=8) 61.8% 0.46T
− SKV (no fast path) 63.1% 0.71T

OK-VQA · Single Component

lower bound
Full SKIP (all components) 63.2% 0.42T
QVS only (others dense) 58.3% 0.95T
RCSR only (others dense) 60.1% 1.05T
BSCA only (others dense) 60.8% 0.88T
Joint sparsity beats individual sparsity by 2–5 points at matched FLOPs

FLOPs Breakdown — LLaVA-1.6-13B Forward Pass

Table 18
Component Dense (GFLOPs) SKIP (GFLOPs) Reduction
Vision encoder72.472.41.0×
QVS scoring0.5+0.3%
RCSR retrieval48.219.32.5×
SKV gate0.8negligible
DBC controller0.3negligible
BSCA (ρᵥL × Δ̄)143.714.69.8×
LLM decoder248.476.43.3×
Total512.7184.32.8×

First non-vacuous sparsity bound for KI-MMQA.

We formalize the intuition that mutual information between visual content and answers concentrates on a small set of salient regions, so the optimal retention rate should be sublinear in V.

Under two mild assumptions — piecewise-Lipschitz mutual information and bounded QVS saliency error η ≤ 0.013 — we prove a closed-form bound. At V = 576 (24×24 patch grid), the bound predicts V′ ≈ 51 retained tokens; the empirical optimum is V′ = 64, consistent within a small constant.

An analogous bound applies to retrieval sparsity: K′ ≥ C′√K log(1/ε), with K′ = 4–8 sufficing in practice at K = 16, ε = 0.05.

ASSUMPTION 4.1 — Piecewise-Lipschitz MI
I(V_S; a | q) is L-Lipschitz in |S|/V on intervals of width ≥ 1/√V. Holds for any smooth feature aggregation function.
ASSUMPTION 4.2 — Bounded Saliency Error
QVS scores satisfy |sᵢ − I(vᵢ; a | q)| ≤ η uniformly. Empirically η ≈ 0.013 on OK-VQA (Spearman ρ = 0.81).
Theorem 4.3 — Sparsity–Accuracy Tradeoff
Under Assumptions 4.1 and 4.2, retaining the top-V′ tokens by QVS score with
V′ ≥ C · √V · log(1/ε)
suffices to ensure I(V_S′; a | q) ≥ I(V; a | q) − ε, where C depends only on L and η.
Corollary 4.4 — For V = 576, retaining V′ = 48–72 tokens (8–12.5%) suffices for ε = 0.05. Matches empirical sweet spot at V′ = 64.
Theorem C.1 — Efficiency–Fidelity Tradeoff
Under Assumptions C.1–C.3:
Acc(SKIP) ≥ Acc(Dense) − O(ε_v + δ_r)
E[FLOPs(SKIP)] ≤ ρ_v C_attn + ρ_r C_ret + O(LΔ̄)
DBC Pareto Optimality (Corollary C.1): The linear budget allocation B*(d) = B_min + (B_max − B_min)·d achieves maximum expected accuracy subject to expected FLOPs ≤ F.
QVS vs. Baselines · OK-VQA Accuracy at Sparsity
65 60 55 50 45 5% 10% 20% 30% 50% QVS (ours) Random DynamicViT 11% = sweet spot

Five complementary axes of knowledge-intensive reasoning.

OK-VQA
Commonsense and world-knowledge questions requiring external knowledge not derivable from image content alone.
14k questions
63.2% SKIP acc.
A-OKVQA
25k questions with reasoning rationales and world-knowledge requirements. Direct-answer and multiple-choice splits.
25k questions
58.4% SKIP acc.
InfoSeek
1.4M entity-centric questions at scale. Target entity often occupies <10% of image area — retrieval quality is primary bottleneck.
1.4M questions
+4.3pt over RA-CM3
Encyclopedic-VQA
Fine-grained categories — birds, butterflies, plants. 1M questions answered with iNaturalist+Wikipedia knowledge base.
1M questions
+3.4pt over RA-CM3
ViQuAE
Entity-grounded questions about named people, places, and works. Uses a 1.7M entity-description KB.
3.7k questions
37.4% F1 score
Memory Footprint
SKIP's BSCA reduces peak KV cache from 12.7 GB (RA-CM3) to 4.8 GB, enabling deployment on consumer 16 GB GPUs.
15.2 GB peak (7B)
62% KV reduction

Honest accounting of the 36.8% SKIP misses on OK-VQA.

36.8% queries missed on OK-VQA
Long-tail entities — poor retrieval coverage
Corpus gap, not architecture deficit
38%
Compositional / multi-hop retrieval
BSCA sparse pattern misses 2nd hop
15%
Counting under occlusion
SKV over-confident fast-paths
12%
Ambiguous questions 9%
Annotator disagreement 8%
Fine-grained visual, OCR, other 18%

SKIP in action: what gets retained, retrieved, fused.

Entity — flag on sailboat stern

✓ Correct
Q: "What is the capital of the country whose flag appears on this boat?"
QVS49/576 tokens retained (8.5%) — 38 on flag region, 11 on hull
RCSR2 regions. Flag → Norwegian flag, Scandinavian maritime symbols. Hull → vessel types (uninformative)
BSCA14/392 edges survived (3.6%), all connecting flag tokens to flag chunks
→ "Oslo" ✓

Perceptual — SKV fast path

Fast Path
Q: "How many pizzas are visible?" (kitchen scene, two pizzas on counter)
SKVDrafter confidence: 0.93 > τ_SKV = 0.82. Fast-path triggered — full pipeline not invoked.
FLOPsRetrieval skipped entirely. RCSR and BSCA not executed.
→ "two" ✓ — at negligible compute cost

Encyclopedic — fine-grained bird

✓ Correct
Q: "What is the typical clutch size of this bird species?"
QVS64/576 tokens (11.1%) — concentrated on plumage pattern (species identifier)
RCSR3 regions. Retrieved ornithology references for identified species from Wikipedia
BSCA33/512 edges survived (6.5%)
→ "3–5 eggs" ✓ (ref: "typically 4")

Failure — long-tail entity

✗ Retrieval miss
Q: "In what year was this monument erected?" (obscure regional landmark)
QVSCorrectly retains monument region (62 tokens)
RCSRRetrieves nearby better-documented landmarks. Correct landmark has only 2 Wikipedia mentions.
DiagnosisRetrieval failure — not vision or fusion failure. Addressed by SKIP-ADR.
→ "1923" ✗ (correct: "1956")

Six staged phases — joint training collapses in 72% of seeds.

Sequential staging avoids gradient conflicts between the discrete pruning signal and the downstream cross-attention loss. Without it, DBC requests maximum budget and SKV never fast-paths.

STAGE I
Backbone Warm-Up
Fine-tune VLM on VQAv2+GQA+TextVQA. Frozen vision encoder. 10k steps.
STAGE II
QVS Warm-Up
Gumbel-softmax τ_G=1.0 fixed. Train QVS scoring heads only. 5k steps.
STAGE III
RCSR+BSCA Training
τ_G cosine-annealed 1.0→0.1. τ_e annealed 0.3→0.6. 15k steps.
STAGE IV
SKV Distillation
Knowledge distillation from dense oracle. All other modules frozen. F1=0.913.
STAGE V
DBC Calibration
Held-out budget-accuracy mapping. Piecewise-linear Φ_B. <2 GPU-hours.
STAGE VI
End-to-End Fine-Tune
All modules unfrozen. τ_G=0.1 frozen. 3k steps. AdamW lr=5e-6.

Global Optimization Settings

OptimizerAdamW β₁=0.9, β₂=0.95, wd=0.05
PrecisionBFloat16 mixed
Hardware8× NVIDIA A100 80GB
Effective batch1,024 (microbatch 4/GPU)
Memory efficiencyDeepSpeed ZeRO-2
LoRArank-64, α=128, dropout=0.05

Submission Hyperparameters (7B config)

Visual retention ρ0.11 (11%)
Per-region top-k′4
BSCA quantile β0.10
SKV threshold τ_SKV0.82
Max regions R8
Retrieval indexFAISS IVF-PQ, 6.5M passages

Cite our work.

If SKIP is useful in your research, please consider citing our ICML 2026 EMM-QA workshop paper.

@inproceedings{mohammad2026skip, title = {Salient Knowledge Pathways: Sparse Cross-Modal Routing for Efficient Knowledge-Intensive Multimodal Question Answering}, author = {Mohammad, Noor Islam S. and Bayaz{\i}t, Ulu\u{g}}, booktitle = {Proceedings of the Efficient Multimodal Question Answering (EMM-QA) Workshop at the 43rd International Conference on Machine Learning}, series = {PMLR}, volume = {306}, year = {2026}, address = {Seoul, South Korea}, note = {Available at \url{https://pmlrbd.github.io/skip/}} }