Why Generate a Click? Plug-in Decision Heads for Web Agents

1 Introduction

• Most generated action strings encode a choice among eleven action types and a few dozen numbered boxes, while only typing or answering text is open ended.Generating the choice token by token requires sequential decoding and scores only the written string, without an explicit distribution over alternative actions.Source: §1 · p. 1
• Plug-in heads read action type, target box, and argument from a learned mix of five layers of a frozen generator, then pass uncertain actions or new text back to that generator.The type head pools the prefill, the target head pools visual tokens inside each numbered box, and the argument head selects a listed option or WRITE slot.Source: §1 · p. 2
• At τ = 0.9, the cascade stays 0.3 points below to 1.4 points above the generator on seven offline test sets while heads decide 12 to 64% of steps; at τ = 0.5, WebArena success reaches 13.8% (4B) and 14.6% (8B).The WebArena baselines are 12.5% and 13.6%, respectively; the cascade generates 3.1 and 3.3 tokens per step instead of 7.4 and 7.5.Source: §1 · p. 2
• Jointly training the heads and generator lowers the generator’s own step success by 0.5 to 4.9 points, outweighing stronger heads on Mind2Web and WebArena.Source: §1 · p. 2

2 Related Work

• CC-Net predicts factored actions with policy heads, MindAct ranks candidate elements before language-model selection, and GUI-Actor uses an attention head over image patches [Humphreys et al., 2022; Deng et al., 2023; Wu et al., 2026].Source: §2 · p. 3
• Prior efficiency methods reduce computation within text generation or across models, but the agent’s decision is still written out as text.Examples include early exits, selective classifiers, cascades, speculative decoding, visual token compression, and a fast-small/slow-large agent pair.Source: §2 · p. 3
• Earlier action heads replace the language model or handle only part of the action, whereas these heads read type, target box, and text option from a frozen generator’s prefill.The WRITE slot and confidence threshold return control to the generator, making the heads and generation two readouts of one forward pass.Source: §2 · p. 3

3.1 Observation and Action Space

• Each step presents a goal, screenshot with numbered candidate boxes, action history, and text options; actions comprise a type, target box, and argument.Some action types need no box, while typing, selecting, and answering require different arguments.Source: §3.1 · p. 3

3.2 Heads on the Generator’s Prefill

• The three heads read a learned mix of generator layers to predict action type, target box, and text option or WRITE slot.The target head pools visual tokens within each numbered box; WRITE returns novel text generation to the generator.Source: §3.2 · pp. 3–4

3.3 Training on a Frozen Generator

• Training first fits the generator on demonstrations, then freezes it and trains only the heads, leaving the generator’s own actions unchanged.The head loss uses WRITE when demonstrated text is not a listed option.Source: §3.3 · p. 4

3.4 Confidence Cascade with a Shared Prefill

• The cascade executes a head-selected action above threshold τ, but hands uncertain steps or WRITE arguments to the generator using the shared prefill cache.The generator also handles answers; head-only steps require no decoding.Source: §3.4 · p. 4

4.1 Data and Backbones

• Qwen3-VL 4B and 8B generators are trained for one epoch on 213k web steps, then heads are trained for one epoch on the same steps with the generator frozen.For WebArena, the generator and heads additionally train for three epochs on replayed expert trajectories from 115 training tasks disjoint from test tasks.Source: §4.1 · pp. 4–5

4.2 Benchmarks and Metrics

• Offline tests cover three M2W splits and four GUIAct sets; online tests run all 119 MiniWoB++ tasks and 174 WebArena test tasks.Offline M2W evaluation uses up to 64 candidates per step, and a step succeeds when the chosen box is gold and the operation, including text, matches.Source: §4.2 · p. 5

4.3 Compared Systems

• The comparisons include generator-only, heads-only, and cascade systems at τ = 0.9 and 0.5, plus a jointly trained adapter used for generation or cascading.At τ = 0.9, the heads decide 12 to 18% of M2W steps and 27 to 64% of GUIAct steps; at τ = 0.5, they decide 37 to 46% and 66 to 89%, respectively.Source: §4.3 · p. 5, Table 1 · p. 5

5.1 Offline Benchmarks

• At τ = 0.9 the cascade tracks its generator on all seven offline test sets while heads decide 12% to 64% of steps.The head-decision share ranges from 12% on M2W cross-website to 64% on GUIAct single, extended.Source: §5.1 · p. 5

5.2 Online Agents on MiniWoB++ and WebArena

• On WebArena the cascade is strongest at both scales, with heads acting alone on 51% (8B) to 52% (4B) of steps and generated tokens below half.The table reports online task success and generated tokens per WebArena step.Source: Table 2 · p. 6, §5.2 · p. 6

5.1 Offline Benchmarks

• At τ = 0.5, M2W accuracy improves by up to 3.1 points for 4B and 1.8 on 8B cross-website, while GUIAct loses 1.0 to 1.9 points at both scales.The threshold moves accuracy in opposite directions across these benchmarks.Source: §5.1 · p. 6

5.2 Online Agents on MiniWoB++ and WebArena

• 41.8% (4B) and 36.9% (8B) WebArena success on 47 seen-template tasks contrasts with below 7% for both systems on 127 new-template tasks.On the seen tasks, the corresponding generators reach 34.8% and 33.3%; on new templates, the cascade trails the 4B generator by 0.8 points and ties the 8B one.Source: §5.2 · p. 6

5.3 Speed and the Confidence Threshold

• The 4B MiniWoB++ cascade runs 1.13 times faster at τ = 0.9 and 1.39 times faster at τ = 0.5, with timing-run success within 0.2 points of the earlier results.Table 4 measures model time per step and the share of steps on which heads act alone.Source: Table 4 · p. 6, §5.3 · p. 6
• On M2W cross-website, τ = 0.5 yields 44.0% step success versus 39.8% under random allocation, while GUIAct multi extended should stay at τ = 0.8 or above.The useful threshold range depends on the test distribution.Source: Figure 3 · p. 6, §5.3 · p. 7

5.2 Online Agents on MiniWoB++ and WebArena

• At τ = 0.9 MiniWoB++ success is 53.1% (4B) and 55.4% (8B), within 0.2 points of the generator; at τ = 0.5 it falls by 0.8 and 1.9 points.Heads act alone on 29% and 33% of steps at τ = 0.9, and 67% and 69% at τ = 0.5.Source: §5.2 · p. 6

5.4 Comparison with Published Systems

• Under the standard Mind2Web protocol, two-model cascades reach 48.8 versus 45.0 on 8B cross-website, but Aguvis 7B leads all tested systems.The passage notes that published systems use their own candidate sets or the full page as reference points.Source: §5.4 · p. 7

6.1 Joint Training Costs Generation

• Joint training lowers generation success by up to 4.9 points across seven offline sets, while stronger heads do not offset the cost on WebArena and five of six M2W columns.On MiniWoB++ the joint cascade reaches 53.2% and 54.2%, versus 52.5% and 53.6% for the authors’ cascade; the two are within 0.8 points on GUIAct.Source: §6.1 · p. 7
• The two-model cascade gives the best M2W results at both scales, but requires a second adapter in memory and a second prefill on handed-back steps.At 8B, its reported results are 51.4, 48.5, and 49.1.Source: §6.1 · p. 7, Table 6 · p. 7

6.2 Where Heads Beat Generation

• On M2W, heads solve their decided steps 8.3 points more often than the generator would, while on GUIAct the generator leads on those steps by 1.4 and 2.9 points.The GUIAct differences, weighted by the 85% and 66% of steps decided by heads, correspond to losses of 1.2 and 1.9 points.Source: §6.2 · p. 7, Table 7 · p. 8
• Typing improves by 9.1, 9.1, and 6.3 points on the three M2W splits, while selecting improves most on unseen websites, from 14.8% to 39.8%.Clicks change by +0.1, +0.2, and −0.8 points while heads decide about a third of them; the argument head scores only the selected box’s choices.Source: §6.2 · pp. 7–8

6.3 Target Head and Layer Weights

• Box pooling leads patch attention on four of five test sets, including 37.3 against 35.3 on M2W cross-task.On M2W, 82% of test steps contain a larger candidate enclosing the gold box, so scoring by attention mass favors that larger box unless the smallest-candidate rule is used.Source: §6.3 · p. 8, Table 8 · p. 8
• The learned layer mix weights layer 24 highest at 0.34 and the last layer lowest at 0.07 on 4B; joint training shifts the largest weight to layer 30 at 0.38.The 8B generator has the same profile as 4B.Source: §6.3 · p. 8

6.4 Heads on a Reasoning Generator

• On MiniWoB++, the reasoning generator trails the generator without thoughts by 4.0 points and runs seven times slower, while joint heads raise success from 51.0 to 51.5% and cut time per step from 1.56 to 0.70 seconds.The 4B reasoning generator was trained on MiniWoB++ only; its thoughts were selected or written for 57% of 7,675 deduplicated steps.Source: §6.4 · p. 8, Table 9 · p. 8

7 Conclusion

• At a strict threshold, the agent keeps generator accuracy; at a looser threshold, it gains on unseen websites and WebArena while generating less than half the tokens.The heads handle only the steps they read well, leaving uncertain steps and new text to the same generator.Source: §7 · p. 8
• The decision heads read decisions from the prefill the web agent already computes.The conclusion describes this as a way to act without requiring the generator to produce the action text at every step.Source: §7 · p. 8
• The cascade reserves generator decoding for uncertain steps and steps requiring new text.The heads decide steps they read well and pass the rest to the generator.Source: §7 · p. 8

Limitations

• Agents acting through raw screen coordinates would need a different target head.The current heads choose among numbered boxes.Source: Limitations · p. 9
• The heads are tied to one generator, so they were trained again after that generator was fine-tuned for WebArena.The experiments use Qwen3-VL backbones, a single action grammar, and a threshold fixed per experiment.Source: Limitations · p. 9
• The experiments cover Qwen3-VL backbones and a single action grammar, with the threshold fixed per experiment.Source: Limitations · p. 9

A Data Conversion

• The converted corpus contains 212,761 training examples after removing 4,472 duplicate and 426 unparsable steps, with 10,430 steps held out for validation.The training set includes repeated MiniWoB++ steps; Table 10 reports source counts before deduplication and excludes held-out MiniWoB++ tasks.Source: A Data Conversion · p. 12, Table 10 · p. 12
• MiniWoB++ uses successful policy rollouts and scripted demonstrations, crops and enlarges the task area, and holds out 12 tasks from ten families.Mind2Web windows are scaled by 0.75; small boxes and large negative containers are dropped, and all positive candidates inside the window are gold boxes.Source: A Data Conversion · p. 12
• GUIAct candidates come from its element list, while UGround boxes and WebLINX visible elements become candidates for their respective goals.UGround goals use referring expressions; WebLINX goals use the instructor’s last two utterances, with demonstrations including clicks, text input, and selections.Source: A Data Conversion · p. 12
• Text options combine goal-derived strings, selected page-element names, and select-element choices, with the list capped at 24 lines.In 7,641 training steps the demonstrated text is absent from the options.Source: A Data Conversion · p. 12

• Frozen heads train with the backbone in inference mode, so only heads keep optimizer state; the generator and joint adapter train with data parallelism on two GPUs.Source: B Training Details · p. 12
• Runs use AdamW; thought-writing generators train for three epochs on MiniWoB++ only and decode at most 160 tokens.Source: B Training Details · p. 12, Table 11 · p. 13
• The cited results report 74.6% for target box and 86.1% for argument; 8B heads reach 92.2%, 75.2%, and 87.2%.Source: B Training Details · p. 13

• Replay converts expert trajectories into the agent’s viewport, candidate boxes, history, and action grammar, yielding 871 training and 45 validation steps after filtering.Source: C WebArena Harness and Expert Replay · p. 13
• Live-site evaluation resets and logs into six sites; actions have a 24-token limit, and environment errors count as failures.Source: C WebArena Harness and Expert Replay · p. 13
• 5.8 (4B) and 7.8 (8B) points: the cascade gains most on Shopping, but loses on Shopping admin.Source: C WebArena Harness and Expert Replay · p. 13, Table 12 · p. 13
• Tables 13 and 14 list offline step-success and online task-success results for each seed.Source: Table 13 · p. 14, Table 14 · p. 14

D Per-Seed Results

• Per-seed tables pair each seed’s generator with its matching heads and use the same online environment seeds and tasks across systems.Source: D Per-Seed Results · p. 14

E Threshold Curves and Operations

• GUIAct hover is one source of cascade loss: it solves 5.4% of 49 multi-step-test hover steps, versus 10.9% for the generator.The type mask permits hover only on menu items, which GUIAct elements never are.Source: E Threshold Curves and Operations · p. 14, Table 16 · p. 15

F Latency Measurement

• 0.414 to 0.364 seconds (4B) and 0.369 to 0.240 (8B): cascade model time per WebArena step falls at both scales.Generated tokens are reported as the WebArena cost because they do not depend on GPU sharing.Source: F Latency Measurement · p. 15

G Mind2Web Standard Protocol

• The Mind2Web standard-protocol table reports element accuracy, operation F1, and step success, counting out-of-window gold elements as failures.Published comparison numbers use other candidate sets or full pages, unlike this paper’s windows and candidate lists.Source: Table 17 · p. 15

H Further Related Work

• The related-work discussion distinguishes probing or editing model knowledge from reading a formed decision, and connects cascade savings to visual-token compression savings.Source: H Further Related Work · pp. 15–16

I Case Studies

• Case studies show varied control allocation: heads handle all six Reddit steps, while the shopping-admin examples mix generator handoffs with head actions or use heads for fourteen clicks.Source: I Case Studies · p. 16