The paper introduces plug‑in decision heads that read hidden states from a frozen vision‑language model to predict web‑agent actions, bypassing the need to generate action‑specific text. These heads determine action type, target element, and text options, and a confidence cascade decides when to execute the head’s prediction versus letting the generator continue. Experiments with Qwen3‑VL show that the heads can decide 12–64 % of steps while keeping performance within 0.3–1.4 points of the full generator, and a looser threshold improves success on unseen sites and lifts WebArena task success with fewer generated tokens.
arXiv:2606. 05597v1 Announce Type: new Abstract: Training vision-language web agents with multi-step RL is compute-intensive, with two dominant forms of inefficiency: idle GPUs in synchronous RL, and trajectories that use more steps and tokens than necessary.
arXiv:2606. 05597v3 Announce Type: replace Abstract: Training vision-language web agents with multi-step RL is compute-intensive, with two dominant forms of inefficiency: idle GPUs in synchronous RL, and trajectories that use more steps and tokens than necessary.
The paper introduces WebMRE, an offline benchmark comprising 541 tasks and 5,293 steps extracted from WebArena trajectories, designed to provide deterministic scoring for web agents without live environments. It enables the first systematic study of how guide sentences and grounded actions reinforce each other, showing that jointly decoding a guide improves element selection accuracy and that the guide acts as a causal instruction channel. The authors fine‑tune models that outperform leading zero‑shot baselines on all offline metrics.
The paper introduces GLANCE, a one‑pass block drafting method that enables lossless speculative decoding for vision‑language models. By using a block‑diffusion head that reads the fused vision‑language state, GLANCE eliminates the need for the drafter to process the image at every step, allowing it to fill an entire block in a single forward pass. Experiments show that GLANCE can decode up to 2.93× faster than autoregressive decoding while maintaining exact greedy decoding results across multiple tasks.
The paper introduces WebLoop, an execution‑grounded framework that jointly trains generation, critique, and refinement within a shared policy for functional web generation. WebLoop trains an execution‑free Critic using requirement‑level discriminability and downstream helpfulness signals, then incorporates consequence‑aware credit to improve diagnosis and repair. Experiments with Qwen3.5‑9B show significant gains on WebRise and WebGen‑Bench, with benefits that transfer to larger models and multimodal settings.
arXiv:2608. 06474v1 Announce Type: new Abstract: Large language models increasingly generate complete websites from natural-language descriptions, and reinforcement learning has become a central approach to closing their remaining functional gap.
arXiv:2607. 12640v1 Announce Type: new Abstract: Reinforcement learning with verifiable rewards, and Group Relative Policy Optimization (GRPO) in particular, is now run routinely on a supervised checkpoint in the hope of producing a stronger agent.
arXiv:2609.36642v1 Announce Type: new Abstract: Language-model agents are usually trained by reinforcement learning from one reward per episode, and privileged self-distillation enriches it by lettin...
LEV is a decision model that performs reasoning by looping its layers instead of generating text, allowing it to deepen its computation on a per-question basis. Trained jointly across all loop steps, a single LEV checkpoint can answer at any depth, improving accuracy—especially on hard questions—while keeping inference fast. On benchmark tests, LEV outperforms dense and cascade models by 5.4 to 8.8 accuracy points with sub‑50 ms latency, and on‑demand depth selection saves 15–24% of computation for minimal accuracy loss.
The paper introduces a denoised counterfactual protocol to accurately measure the selectable headroom of world action models (WAMs) and uses these measurements to build a fail‑gated selector. Across four WAMs and two benchmarks, the study shows that the standard estimate overstates selectable headroom, that structured action perturbations can add up to 16 points of headroom, and that imagined futures retain candidate‑outcome information without loss. The selector improves closed‑loop task success from 76.9 % to 81.6 % on held‑out training episodes and from 62.6 % to 69.8 % on unseen tasks.
arXiv:2610.03036v1 Announce Type: cross Abstract: We present WebFovea, a vision-based web agent that placed 2nd in the WebRetriever Challenge 2026 with a final score of 57.0 out of 100. The challenge...
The paper critiques current GUI grounding benchmarks that count a click as correct only when it falls inside a single annotated box, arguing that many interfaces expose the same action through multiple elements. The authors propose redefining correctness as achieving the same outcome as clicking the annotated target, and introduce a probe that executes and fingerprints thousands of clicks per page to create outcome maps and an execution judge. Applying this method to several large‑scale grounders on real web applications shows that a significant fraction of clicks rejected by box‑based scoring actually produce the intended outcome, indicating that box scoring under‑counts correct clicks and that evaluation metrics need adjustment.