{"id":"708774d9-8e2a-4462-b80a-a7ac5d23ed83","arxiv_id":"2606.09399","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"SUPERBROWSER reaches 89.47% success on Mind2Web Hard by implementing a human-like perception-cognition-action system with vision-first bounding boxes, a three-role brain, and an evicting ledger.","lead":"SUPERBROWSER is an autonomous web agent that copies human browsing by using vision to spot targets first, splitting reasoning into three roles, and keeping only a short memory ledger that forgets old details. A smart generalist might read it to see how copying human limits on attention and memory can make AI agents better at real web tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Performance attribution to cognitive contract lacks ablations or VLM-controlled baselines","rationale":"The reader's weakest assumption directly identifies the same attribution gap. Full text does not add the missing controls or VLM details needed to close it, so the concern remains load-bearing and the UNVERDICTED stance is appropriate.","tokens_in":1816,"tokens_out":292,"duration_ms":12228,"concrete_test":"Reproduce the 66 Mind2Web Hard tasks with the identical VLM and action executor but replace the three-role brain + Ledger with a single ReAct loop that receives the same bounding-box observations; if success falls below ~70% the design contribution is supported, otherwise the attribution weakens.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline result (89.47% on Mind2Web Hard) is presented as evidence that consistent application of the perception-cognition-action triad drives gains over open baselines. This requires that the three mechanisms (vision-first bounding boxes, Orchestrator/Planner/Worker, Ledger with eviction) are causally responsible rather than the underlying VLM, prompt engineering, the three-tier click cascade, or benchmark-specific tuning. The manuscript supplies no ablation removing individual components while holding the VLM fixed, no comparison against a ReAct-style agent on the identical model, and no disclosure of the exact VLM or its zero-shot capability on the same tasks.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper presents SUPERBROWSER, an autonomous web-navigation agent designed to emulate human browsing via a perception-cognition-action triad. This is implemented through a vision-first bounding-box pipeline for candidate regions, a three-role brain (Orchestrator for routing, Planner for progress evaluation, Worker for actions), and a structured Ledger with six-phase eviction to retain only goal-relevant information. On the Mind2Web Hard benchmark (66 tasks), it reports 89.47% success, ranking third overall and ahead of published open/research baselines, with gains attributed to consistent application of the cognitive contract rather than isolated tricks.","tokens_in":1948,"tokens_out":450,"duration_ms":21773,"significance":"If the performance gains can be causally linked to the proposed mechanisms through controlled experiments, the work would contribute a human-grounded design principle for web agents that could improve reliability and reduce context bloat in long-horizon tasks. The high success rate on a challenging benchmark is notable and, if substantiated, offers a falsifiable template for agent architecture that other systems could adopt or test against.","major_comments":[{"comment":"Abstract: The central claim that 'the gain comes not from any single trick but from the consistent application of a cognitive contract throughout the system' is load-bearing for the paper's thesis. No ablation results, component-removal experiments, or VLM-controlled baselines (e.g., ReAct-style agent on the identical model) are referenced to isolate the contribution of the bounding-box pipeline, three-role separation, or Ledger eviction; the attribution therefore remains an assertion rather than a demonstrated result.","section":"Abstract"},{"comment":"Experimental evaluation (implied by benchmark reporting): The manuscript does not disclose the exact VLM employed or its zero-shot performance on the same Mind2Web Hard tasks. Without this, it is impossible to determine whether the 89.47% success rate exceeds what the underlying model would achieve under a simpler prompting regime, undermining the claim that the cognitive contract is the source of the margin over baselines.","section":"Experimental evaluation"}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for these focused comments on evidence and attribution. We address each point directly and commit to revisions that add the requested controls and disclosures without altering the core claims.","responses":[{"response":"We agree the claim requires stronger isolation. The revised manuscript will add (i) component-removal ablations for the vision-first bounding-box pipeline, three-role brain, and six-phase Ledger eviction, and (ii) a ReAct-style baseline run on the identical VLM and task set. These experiments will quantify the incremental contribution of the integrated cognitive contract versus isolated mechanisms.","revision_made":"yes","referee_comment":"[Abstract] Abstract: The central claim that 'the gain comes not from any single trick but from the consistent application of a cognitive contract throughout the system' is load-bearing for the paper's thesis. No ablation results, component-removal experiments, or VLM-controlled baselines (e.g., ReAct-style agent on the identical model) are referenced to isolate the contribution of the bounding-box pipeline, three-role separation, or Ledger eviction; the attribution therefore remains an assertion rather than a demonstrated result."},{"response":"We will name the exact VLM and its version in the revised experimental section. We will also report zero-shot and ReAct-style prompting results on the identical model and Mind2Web Hard split so readers can directly compare the margin attributable to the proposed architecture.","revision_made":"yes","referee_comment":"[Experimental evaluation] Experimental evaluation (implied by benchmark reporting): The manuscript does not disclose the exact VLM employed or its zero-shot performance on the same Mind2Web Hard tasks. Without this, it is impossible to determine whether the 89.47% success rate exceeds what the underlying model would achieve under a simpler prompting regime, undermining the claim that the cognitive contract is the source of the margin over baselines."}],"tokens_in":1538,"tokens_out":411,"duration_ms":20466,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The paper introduces SuperBrowser, a web agent built around the idea that agents should browse like people: scan a few targets, decide, and keep only the necessary state. It implements this with a vision-first bounding-box pipeline, a three-role brain (orchestrator, planner, worker), and a ledger that uses a six-phase eviction loop to drop stale data. On the 66-task Mind2Web Hard set it reaches 89.47% success and beats published open baselines.\n\nThe concrete mechanisms are new in combination: the asynchronous prefetch of bounding boxes, the separation of strategic and operational roles, and the explicit eviction rules for screenshots and traces. The action layer with its three-tier click cascade and Bezier motion is also described in usable detail.\n\nThe main weakness is the missing link between the design and the result. The abstract asserts that consistent application of the cognitive contract produces the gain, yet supplies no ablation that removes one piece while holding the VLM fixed, no ReAct-style baseline on the same model, and no statement of which VLM was used or what its zero-shot score would be. Without those, the performance number stands but the causal story does not.\n\nThis is for people building practical browser agents who want an architecture sketch they can try. It is worth sending to peer review because the system is fully specified and the benchmark number is competitive, even though the attribution claim will need tighter evidence.","headline":"SuperBrowser reports 89% on Mind2Web Hard with a human-inspired three-role brain and ledger, but the gains are not tied to those components by any controls.","tokens_in":2402,"tokens_out":368,"would_cite":false,"duration_ms":16863,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Web agents reach high success when they limit memory and separate perception from planning like humans do.","keywords":["autonomous web navigation","human browsing behaviour","cognitive contract","vision-first pipeline","structured ledger","browser agent","mind2web benchmark","perception-cognition-action"],"falsifier":"An ablation that keeps the same model and infrastructure but removes the Ledger eviction loop or the three-role separation and measures whether success on the same 66 tasks falls below the reported 89.47 percent.","tokens_in":2719,"feed_emoji":"🧭","tokens_out":792,"duration_ms":18274,"temperature":0.7,"pith_summary":"The paper claims that an autonomous web-navigation agent succeeds when it copies the way a person actually browses: looking at a few candidate targets first, deciding, and retaining only the minimal information needed to stay on goal. It implements this through a vision-first pipeline that identifies interactive regions on screenshots before the language model sees them, a three-role brain that routes tasks, checks progress, and issues actions separately, and a structured ledger that keeps the goal, recent actions, key facts, and checkpoints while evicting everything else through a fixed six-phase loop. On the Mind2Web Hard benchmark of 66 tasks the resulting system records 89.47 percent success, third overall and ahead of all published open browser-agent baselines. The authors attribute the margin to the uniform application of these human-style constraints rather than to any single added feature.","feed_headline":"Human-mimicking web agent hits 89% on hard tasks","feed_subtitle":"Limiting memory to recent actions and facts while separating perception from planning outperforms open baselines on Mind2Web Hard.","key_machinery":"The perception-cognition-action triad: vision-first bounding-box pipeline, three-role brain, and eviction-managed Ledger that retains only goal-relevant state.","core_discovery":"SUPERBROWSER operationalizes human browsing as three coupled mechanisms. A vision-first bounding-box pipeline labels candidate interactive regions on every screenshot and supplies them asynchronously to the language model. A three-role brain—an Orchestrator for classification and routing, a Planner that evaluates progress every few steps, and a Worker that emits actions—separates strategic from operational reasoning. A structured Ledger stores only the goal, the last three actions, a small set of facts and dead-ends, and checkpoints; a six-phase eviction loop discards stale screenshots, state blobs, and traces. Action execution uses a three-tier click cascade with humanized Bezier motion and","pith_inferences":["If memory limits are the main driver, the same eviction discipline could improve performance in other long-horizon agent domains such as robotic manipulation.","Removing the vision-first step while keeping the rest would test whether perception ordering matters more than the ledger itself.","The design implies that simply enlarging context windows is not the only or best route to better web agents.","Future work could collect additional human browsing traces to refine the exact eviction rules or role boundaries."],"forward_implications":["Agents achieve high task completion without retaining every pixel or full reasoning trace in context.","Separating progress evaluation from action emission reduces drift over long navigation sequences.","Humanized execution mechanics resolve common UI ambiguities such as small arrows next to labels.","Consistent application of limited memory across components yields larger gains than isolated optimizations.","The same cognitive-contract pattern can be applied to other sequential interaction benchmarks."],"fun_headline_variants":["SuperBrowser mimics human browsing for 89% on Mind2Web Hard","Vision bounding box pipeline precedes actions in web navigation agent","Orchestrator planner worker brain separates reasoning in SuperBrowser","Structured ledger with eviction loop limits memory to recent facts"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The performance gain is produced by the human-like perception-cognition-action design rather than by unstated implementation choices or the underlying model.","fun_headline_variants_meta":{"raw":{"variants":["SuperBrowser mimics human browsing for 89% on Mind2Web Hard","Vision bounding box pipeline precedes actions in web navigation agent","Orchestrator planner worker brain separates reasoning in SuperBrowser","Structured ledger with eviction loop limits memory to recent facts"]},"model":"grok-4.3","cost_usd":0.007585,"raw_usage":{"total_tokens":3547,"prompt_tokens":811,"num_sources_used":0,"completion_tokens":66,"cost_in_usd_ticks":75849500,"prompt_tokens_details":{"text_tokens":811,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2670,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":811,"tokens_out":66,"duration_ms":16845,"temperature":1.0,"reasoning_tokens":2670,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-27T16:35:54.518398+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"An ablation that keeps the same model and infrastructure but removes the Ledger eviction loop or the three-role separation and measures whether success on the same 66 tasks falls below the reported 89.47 percent.","supporting_citations":[],"review_version":1}