⚠️ This repository contains three unrelated experiments — do not mix them

Location	Experiment	Model	Tasks	Methods compared
repository root (dense/ quest/ tsa_*/)	described in this README, below	Qwen3-VL-32B-Instruct (vision off)	WebArena 0–99 (100)	dense / quest / TSA ×3
browseruse_qwen3-32b_sparsity_2026-09/	block-sparse vs quest sparsity sweep	Qwen3-32B	the 608 non-GitLab tasks (complete: 15 × 608 = 9,120 runs)	dense + block sparse ×7 levels + quest ×7 levels
gaia-web_qwen3-32b_sparsity_2026-09/	⚠️ not WebArena — GAIA-web, different harness, LLM-judged	Qwen3-32B	GAIA-web, 90 live-web tasks (9 × 90 = 810 runs)	dense + block sparse ×4 levels + quest ×4 levels

The WebArena experiments differ in model, task set, judge model, sampling temperature and site backends, so pass rates are not directly comparable across them. The GAIA-web experiment is a different benchmark altogether (live web, a different agent harness, success judged by an LLM) and must not be compared with either WebArena experiment. Note also that dense/ and quest/ at the root are not the same data as the dense and quest_* configurations inside the new directory.

The new experiment additionally records GPU energy and thermal telemetry. See browseruse_qwen3-32b_sparsity_2026-09/README.md.

Method configurations at a glance

Root dataset — Qwen3-VL-32B-Instruct, temperature 0, max_tokens 4096, browser-use max_steps 30, judge Llama-3.3-70B-Instruct, 100 tasks per method:

Method	top_k	Selected tokens	Pass rate (lenient, 100 tasks)	Pass rate (official)
dense	—	100%	33%	23%
quest	61 pages	10.8%	33%	26%
TSA-minmax tk64	64 chunks	78.3%	34%	19%
TSA-centroid tk64	64 chunks	57.5%	28%	14%
TSA-centroid tk32	32 chunks	26.6%	26%	9%

New directory — Qwen3-32B, temperature 0.2, max_tokens 8192, thinking on, browser-use max_steps 30, self-judged, the 608 non-GitLab WebArena tasks. Budget: max(topk_val + bos + eos, int(cached_block_len * topk_ratio)), topk_val = 1 (inert at these context lengths), all 64 layers sparse; block sparse uses bos=1, eos=1, quest uses bos=0, eos=1.

Release status: complete (2026-09-16; 9,120 runs = 15 configurations × 608 tasks). Only 15 runs remain defective (dense 6, quest_1p5625 9); they are excluded from the "clean" column.

Configuration	Kernel module	topk_ratio	Runs (of 608)	Pass rate (as scored)	Pass rate (clean runs)
dense	— (sparsity disabled)	—	608	19.2%	19.4%
block_50	gqa_block_sparse_attention	0.5	608	18.8%	18.8%
block_37p5	gqa_block_sparse_attention	0.375	608	17.4%	17.4%
block_25	gqa_block_sparse_attention	0.25	608	18.1%	18.1%
block_12p5	gqa_block_sparse_attention	0.125	608	20.6%	20.6%
block_6p25	gqa_block_sparse_attention	0.0625	608	18.4%	18.4%
block_3p125	gqa_block_sparse_attention	0.03125	608	16.0%	16.0%
block_1p5625	gqa_block_sparse_attention	0.015625	608	11.2%	11.2%
quest_50	gqa_quest_sparse_attention	0.5	608	19.9%	19.9%
quest_37p5	gqa_quest_sparse_attention	0.375	608	20.9%	20.9%
quest_25	gqa_quest_sparse_attention	0.25	608	19.9%	19.9%
quest_12p5	gqa_quest_sparse_attention	0.125	608	15.0%	15.0%
quest_6p25	gqa_quest_sparse_attention	0.0625	608	12.8%	12.8%
quest_3p125	gqa_quest_sparse_attention	0.03125	608	6.6%	6.6%
quest_1p5625	gqa_quest_sparse_attention	0.015625	608	5.6%	5.7%

Paired tests on tasks clean in both configurations (results/paired_tests_clean.json):

• At ≥ 25% sparsity, block sparse, quest and dense are statistically indistinguishable.
• Block sparse holds dense-level accuracy down to 3.125% (−3.3 pp, p = 0.072) and breaks at 1.5625% (−8.1 pp, p < 0.001).
• Quest falls below dense from 12.5% on: −4.3 pp at 12.5% (p = 0.013), −6.5 pp at 6.25%, −12.8 pp at 3.125% and −13.9 pp at 1.5625% (all p < 0.001).
• Block beats quest at every level from 12.5% down (p = 0.0022, 0.0019, < 0.001, < 0.001). At 37.5% quest is ahead (p = 0.031), but that is borderline given seven levels were tested.
• fuzzy_match is judged by the model under test, which itself degrades at the most aggressive sparsity levels; re-judging with a fixed external judge is recommended.

See the sub-directory README for the defect accounting, energy telemetry and harness fixes. fuzzy_match has since been re-judged with two fixed external judges (Llama-3.3-70B-Instruct and dense Qwen3-32B): under them the block-over-quest advantage at 12.5% and 6.25% disappears and block sparse falls below dense from 3.125%. See the sub-directory's "External-judge re-scoring" section and results/rejudge_external/.

GAIA-web experiment (gaia-web_qwen3-32b_sparsity_2026-09/) — a separate benchmark

Qwen3-32B on the 90 GAIA-web tasks (live web, Level 1 and 2), run with a colleague's simulator browser-use harness (headed browser, text-only agent) and judged by the dense model with the WebVoyager success protocol. Single run per configuration.

Configuration	dense	block 25%	quest 25%	block 12.5%	quest 12.5%	block 6.25%	quest 6.25%	block 3.125%	quest 3.125%
Success rate (of 90)	20.0%	10.0%	20.0%	18.9%	21.1%	21.1%	17.8%	11.1%	15.6%

With 90 tasks and one run each, no difference survives correction for multiple comparisons; these numbers are not comparable to the GAIA leaderboard. See its README before using it.

WebArena 0-99 · Sparse-Attention Method Comparison · LLM Trajectories

Full LLM call trajectories, final answers and scores for five attention implementations, each run over WebArena tasks 0-99 (100 tasks per method).

Experiment configuration

Item	Value
Model under test	Qwen3-VL-32B-Instruct (same model for all five methods; only the attention implementation changes)
Scoring judge	Llama-3.3-70B-Instruct
Vision / screenshots	Disabled throughout (use_vision=False + chromium imagesEnabled=false); the agent only receives the page's accessibility tree as text
Sampling	temperature 0, max_tokens 4096
Agent	browser-use, max_steps 30
Map site	points at the real openstreetmap.org (not a self-hosted snapshot)

Note: although the model is a vision-language model, vision was deliberately turned off — vortex/quest does not handle mrope and produces garbage with images, and disabling it keeps the input modality identical across all five methods. All numbers here are therefore text-only agent results and do not represent the model's ceiling with vision enabled.

Results overview (lenient is the primary metric)

Method	top_k	selected tokens	non-map (48 tasks)	TOTAL (100)	official
dense	—	100%	19/48 (39%)	33/100	23%
quest	61 pages	10.8%	19/48 (39%)	33/100	26%
TSA-minmax tk64	64 chunks	78.3%	17/48 (35%)	34/100	19%
TSA-centroid tk64	64 chunks	57.5%	14/48 (29%)	28/100	14%
TSA-centroid tk32	32 chunks	26.6%	11/48 (22%)	26/100	9%

selected tokens is the measured fraction of the KV cache the method actually attends to (full-model measurement with real decode queries). It is the only directly comparable axis between quest and TSA — %chunks and %pages are different units.

Directory layout

<method>/
├── SCORES.json            official scoring summary
├── SCORES_adjusted.json   lenient scoring summary
└── task_<id>/
    ├── llm_calls.jsonl    ★ full LLM call trajectory (one JSON object per line)
    ├── task_<id>.json     final answer, step count, final_url, timings
    ├── input.json         resolved task definition (intent / eval / replica map)
    └── run.log            browser-use agent log
task_ids/                  official & lenient pass/fail task-id lists per method

method ∈ dense · quest · tsa_minmax_tk64 · tsa_centroid_tk64 · tsa_centroid_tk32

Field reference

llm_calls.jsonl — one LLM call per line

Field	Description
call	call index within the task
t / latency_s	start timestamp / call latency in seconds
input_messages	full input messages (system preamble + page accessibility tree)
output	model output (browser-use structured action JSON)
usage	token counts. ⚠️ the TSA server does not report prompt_tokens (always 0) — re-tokenize if you need input length; completion_tokens is valid for all methods

task_<id>.json — intent, answer, final_url, n_steps, is_done, wall_time_s, steps (per-step url / next_goal / actions), error (TimeoutError means the task timed out).

input.json — intent, eval (eval_types + reference_answers, the scoring ground truth), sites, replica_map (which site replica this task was assigned to).

Caveats

• official vs lenient: official is the standard WebArena evaluator; lenient additionally uses an LLM judge to recover answers that state the correct value but fail strict string matching on phrasing/formatting. Map tasks run against the real OSM while reference answers were annotated on a self-hosted snapshot, so official is systematically low — use lenient as the primary metric.
• ~25% of TSA tasks time out (error: TimeoutError), mostly map tasks. The TSA server has no tensor-parallel support and runs the 32B model on a single GPU, making it ~4× slower per request than the tp=4 dense/quest servers; its task timeout was raised to 2400s to compensate.
• Do not draw speed/throughput conclusions from this data: dense ran with --enforce-eager (CUDA graph disabled) while quest/TSA had it enabled, and parallelism differs (single-GPU vs tp=4). Accuracy conclusions are unaffected.
• Site URLs in the trajectories point at an internal experiment environment and are only meaningful there.

Reproduction code and full methodology (harness, scoring scripts, sm90 build patches, pitfalls checklist) live in the companion reproduce/ directory.