shiqihe/WebVoyager-GAIA-SparseAttention-Offline-Qwen3-VL-30B-A3B

Sparse Attention for Web Agents — Offline Replay Records

Step-level records from an offline replay study comparing three sparse-attention methods against full attention on browser-agent trajectories.

Model: Qwen3-VL-30B-A3B-Instruct (48 layers, 128 experts / 8 active, GQA 32:4, head_dim 128) Hardware: NVIDIA GB10 (sm_121) Tasks: 50 complete trajectories sampled from WebVoyager + GAIA — 332 steps, of which 190 emit an element index. Every task was completed successfully by the reference model (Qwen3.5-Omni, dense attention), so every step has a reference action.

The comparison

All three methods chunk the KV cache, score chunks against the current query, keep the top-scoring ones, and attend only to those. They differ in exactly two design choices:

Method	Chunking	Scoring
TSA	variable, aligned to document structure (16–256 tokens)	envelope (per-dimension min/max upper bound)
Quest	fixed, 16 tokens	envelope
BlockSparse	fixed, 16 tokens	centroid (mean of chunk keys)

TSA differs from Quest only in chunking; Quest differs from BlockSparse only in scoring, so the comparison isolates one variable at a time.

Budget = page_size × top_k — the number of KV tokens actually attended per decode step. This is the fairness control: top_k alone is not comparable across methods because their page sizes differ 4x. For scale, at ~29.5k tokens of context a 4096 budget reads about 1/7 of it.

Files

Each data/<method>_budget<N>.jsonl holds 332 step records for the same 50 tasks.

Field	Meaning
task	task identifier, e.g. Allrecipes--35
site	site the task runs on
step	step index within the trajectory, e.g. step_001
is_idx	whether this step's action selects a page element by index
chosen	element index/indices the model produced
ref	element index/indices in the reference trajectory
types	action type(s) produced, e.g. ['input']
ref_act	reference action type
verdict	valid (index exists on the page), none (no index emitted), otherwise invalid
agree	whether chosen matches ref exactly
head	leading text of the model's output
err	error, if the step failed

data/replay_contexts.jsonl — the 332 replayed contexts (the exact prompt each config saw). data/task_ids.json — the 50 task identifiers. data/summary_h1_*.json — per-config aggregate counts.

Headline results

Index-step agreement with the reference trajectory (190 index steps):

Config	Budget	valid	agree
dense_full_attention	—	81.1%	51.1%
tsa_budget4096	4096	70.5%	28.9%
quest_budget4096	4096	69.5%	20.5%
blocksparse_budget4096	4096	71.6%	18.4%
tsa_budget8192	8192	79.5%	43.7%
quest_budget8192	8192	73.2%	47.4%
blocksparse_budget8192	8192	79.5%	47.9%
tsa_budget16384	16384	79.5%	51.1%

Three things worth knowing before using these numbers:

1. agree has a ceiling well below 100%. Full attention itself scores 51.1%; 30% of its steps pick a valid but different element. Many tasks have several correct paths, so disagreement is not the same as error. Judge sparse configs against ~51%, not 100%.
2. The ranking flips with budget. At 4096 TSA is significantly better than both baselines (cluster-level sign test, p = 0.015 / 0.020). At 8192 the baselines catch up and TSA is the only method still significantly behind full attention (p = 0.0042). At 16384 TSA reproduces full attention exactly. Quest and BlockSparse were not run at 16384, so no ranking above 8192 is supported either way.
3. Sparsity causes index hallucination. Full attention emitted an index that does not exist on the page 0 times out of 190; sparse configs do it 1–10% of the time. Conversely, "emitted no index at all" sits at 18–24% for every config including dense, so that failure is not attributable to sparsity.

Caveats

• These are offline replay records: the reference context is replayed at every step, so errors do not compound as they would in a live agent loop. They are a proxy for task success, not task success itself.
• Statistics should be cluster-level with the task as the unit — steps are nested within tasks, and treating 190 steps as independent overstates significance.
• quest_budget4096 was originally run over 100 tasks; the rows here are the 50-task subset shared by every other config.