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. We ask whether it adds skill to a small language and vision-language model web agent at the 4B to 8B scale, or whether it mostly reshapes behavior the supervised model already has. Across a control grid of 18 runs that varies learning rate, KL weight, seed, initialization, and clipping, no configuration credibly improves the success rate of a strong supervised baseline on tasks the agent has largely mastered. On the text track, moderate to high learning rates make it credibly worse. The null holds under paired testing, 25 evaluation seeds, 6 training seeds, changes to the recipe, both text and Set-of-Marks screenshot observations, and scaling the backbone to 8B; the credible harm is a text-track finding and is only nominal under Set-of-Marks. To show that the null reflects the setting and not a broken pipeline, we run the identical harness, reward, and recipe on tasks whose reward is reachable by sampling, and there the success rate rises by 22 points with a paired interval that excludes zero. GRPO therefore helps only when there is headroom to climb, meaning the sampled policy already succeeds more often than the greedy one. We then explain the failure. A middle learning rate degrades the agent and a high one collapses it, and the two regimes form a double dissociation: grafting localizes the degrade regime to the attention and MLP blocks, while the collapse regime cannot be traced to any single group, and the embedding change that dominates the weight movement is causally inert. At 4B, effective rank in the late layers tracks capability in both directions; at 8B the two come apart. This coupling is specific to the smaller model, so we report it as scale-dependent.
Chengguang Gan, Zhixi Cai, Yunhao Liang et al.· 0 citations
Public tests are widely used to guide large language model code generation, but whether models treat them as executable specifications or merely as extra prompt context remains unclear. We study test-driven code generation on HumanEval+, MBPP+, and recent LiveCodeBench tasks using Qwen2.5-Coder-7B and Qwen3.6-27B. We compare natural-language-only prompts with relevant visible tests, shuffled outputs, irrelevant tests, assertion-only tests, and stronger-model-generated synthetic tests. Evaluation combines hidden or private test pass rates with task-level behavior flips, linear probes, and layer-wise hidden-state shifts. Visible tests substantially improve Qwen2.5 performance on MBPP+ but have little or unstable effect on HumanEval+ and LiveCodeBench. For Qwen3.6 on LiveCodeBench, the natural-language-only pass rate rises from 13.1% to 39.4%, yet relevant original tests add only 2.9 percentage points (p = .458), and synthetic high5 tests add 1.7 points over their matched baseline (p = .701). More tests produce larger representation shifts, but accuracy gains quickly saturate or disappear; assertion-only and irrelevant-test prompts can strongly alter hidden states without improving correctness. After controlling for prompt length, task-level shifts are associated with both beneficial and harmful flips, with a substantially stronger association for harms. These results show that tests influence code models through both semantic guidance and prompt-context perturbation, and that representational change alone does not demonstrate effective test utilization.
Yunhao Liang, Chengguang Gan, Ruixuan Ying et al.· 0 citations
Digital Adoption Platforms (DAPs) are embedded overlays widely used on web systems to guide users through operations inside a page, helping them get started with unfamiliar interfaces quickly. Completing a real task, however, rarely means clicking a few buttons on a single page: it takes a sequence of actions that unfolds across changing page states. Prior studies have also treated automated web agent actions and guide text generation as two separate problems, and most of them feed models textual page representations such as the DOM or accessibility trees rather than the rendered screens that humans actually operate on. In this work we introduce MAG, the first benchmark that unifies task execution and guide writing into a single Multimodal Action and Guide task, with two grounding schemes over screenshots: Set-of-Mark element selection and raw pixel coordinates. We further build a complete harness for this compound task, covering annotation with LLM assistance and human verification, training, evaluation in live environments, and joint metrics for actions and guides. With this harness we evaluate frontier API models and open multimodal models, and report detailed analyses. Finally, we design a GRPO training method augmented with expert trajectories, which nearly doubles the success rate of a supervised 9B agent (from 6.9% to 13.2%) and improves guide quality at the same time. Even the strongest model completes fewer than 40% of the tasks, leaving ample room for future research.
Chengguang Gan, Hanjun Wei, Yunhao Liang et al.· 0 citations
The results show that executable feedback can repair secure-code generation, but its benefits depend on the model, task, feedback entry point, and especially test coverage.
Yunhao Liang, Chengguang Gan, Ruixuan Ying et al.· 0 citations
An audit-and-placebo protocol is proposed that separates verifier artifacts, interaction scaffolding, and grounded feedback credit in evaluations of self-evolving test generators in evaluations of self-evolving test generators.
Yunhao Liang, Chengguang Gan, Ruixuan Ying et al.· 0 citations