Each section below summarizes one paper's reported method, experiments, and limits from its full review. The studies use different tasks and metrics; their results are not directly comparable.
arXiv:2611.07408
PSGAgent: Enhancing MLLMs' Dynamic Visual Perception for Panoptic Scene Graph Generation via Agentic Reasoning
PSGAgent reframes panoptic scene graph generation as an iterative think–act–observe problem for multimodal language models. A 3B Qwen2.5-VL backbone can invoke a Region Zoom-In Tool for small or peripheral relations and a Visual Highlight Tool for ambiguous subject–object interactions, then refine predictions through cold-start supervised fine-tuning on 8K error-aware M-CoT trajectories followed by GSPO reinforcement learning with hierarchical rewards. On the PSG benchmark, the model reports 50.27 Recall@100, 42.88 mean Recall@100, and 46.58 Mean under SGDet, while reaching 73.04 Recall@100, 67.75 mean Recall@100, and 70.40 Mean under PredCls.
How it works
The agent models a graph G = (O, R), where each object contains a category and a non-overlapping binary mask, and each relation is a predicate between a subject and an object. Its two visual operators have complementary purposes: cropping increases resolution for small or boundary relations, while highlighting removes unrelated context around an interacting pair. Error-aware data mining begins with mispredicted triplets from a vanilla MLLM.
What the paper evaluates
The evaluation uses the PSG dataset, which contains 48,749 images, 133 object categories, and 56 predicate categories with panoptic segmentation masks. SGDet requires object detection, category prediction, and relation prediction from the image; PredCls supplies ground-truth object boxes and labels to isolate predicate reasoning. The paper reports Recall@100, mean Recall@100, their average Mean, AP@50, zero-shot Recall@100, and open-vocabulary relation scores under a 7:3 base–novel split. Under SGDet, the authors report 49.91 AP@50, 50.27 Recall, 42.88 mRecall, and 46.58 Mean for PSGAgent.
Limit to keep in view
The experiments are centered on the PSG dataset; the paper does not provide evidence for transfer to video, embodied interaction, or other scene-graph datasets.
arXiv:2610.39324
MRC-MMAgent: Reinforcement Learning for Multi-Role Collaboration in Multimodal Tool-Using Agents
MRC-MMAgent trains a shared multimodal tool-using policy to solve, inspect, and continue within one executable loop. A solver gathers evidence with reverse image search, text search, and image zoom, an inspector judges whether the trajectory sufficiently supports the candidate answer without seeing the reference answer, and a continuation solver revises the solution when evidence is inadequate. A role-coupled reward combines final accuracy, interaction validity, and successful correction, while full-trajectory GRPO updates decisions from all roles. Using only the MMSearch-R1 training set, the method improves the 7B baseline from 51.1 to 62.2 average accuracy and reaches 68.1 average accuracy at 8B across five search-intensive benchmarks.
How it works
The system uses a shared policy conditioned by three frozen role prompts. At loop index k, the initial or continuation solver produces X_k, containing reasoning, actions, observations, and a candidate answer. Equation 1 formalizes this segment, while Equation 2 selects S_1 for the initial pass and S_2 for later passes. The inspector receives the accumulated history h'_k under S_3 and emits C_k, including a rationale and a control decision.
What the paper evaluates
The evaluation covers five knowledge-intensive multimodal search and visual-question-answering benchmarks—FVQA-Test, InfoSeek, SimpleVQA, MMSearch, and LiveVQA—and three visual grounding or high-resolution settings: V* Bench, HR-Bench 4K, and HR-Bench 8K. Training uses Qwen2.5-VL-7B-Instruct and Qwen3-VL-8B-Instruct, reverse image search, text search, and image zoom. The reported setup uses 9 NVIDIA H200 GPUs, batch size 128, learning rate 5e-6, 16K-token sequences, rollout groups of 8, and 10-turn tool interactions, with strict train-test separation.
Limit to keep in view
The main-body evaluation covers Qwen2.5-VL-7B-Instruct and Qwen3-VL-8B-Instruct; transfer to substantially different architectures and model scales is not established.
arXiv:2611.03634
WebHarbor: Docking Real Websites for Evolving GUI Agent Environments
WebHarbor presents 153 executable, Dockerized reconstructions of real-world websites and 3,518 tasks for evaluating and training GUI web agents in controlled offline settings. Tasks cover information seeking as well as authentication, bookings, purchases, account management, and persistent state changes. The authors build environments by using tasks to scope coding-agent implementations, then asking human contributors to audit visual fidelity, functional correctness, and task validity. In experiments, a Qwen 4B-level agent trained with reinforcement learning in WebHarbor transfers to held-out tasks and performs competitively on live-web benchmarks.
How it works
WebHarbor begins with tasks rather than attempting to reproduce an entire commercial service. A task specifies entry and terminal pages, observable entities and attributes, required browser actions, and state changes. These requirements define the routes, data, assets, and interactions the local environment must support. The authors start with the 15 sites and 643 tasks in WebVoyager, then extend site coverage to Online-Mind2Web websites while excluding its original tasks, which they characterize as primarily information-seeking.
What the paper evaluates
The authors evaluate WebHarbor on four benchmarks: its 306-task test set, WebVoyager, Online-Mind2Web, and WebArena. Online-Mind2Web covers 300 tasks across 100+ live websites; WebVoyager has 643 tasks on 15 live sites; WebArena uses a 226-task subset. WebArena-NL disables automatic login and supplies credentials in task descriptions. The protocol uses a 1280×1000 viewport, temperature 0.6, top-p 0.95, and top-k 20, with a 30-step budget except for 50 steps on WebArena. Online-Mind2Web uses an o4-mini AgentTrek judge, WebVoyager uses a GPT-4.1 benchmark judge, and WebArena uses programmatic state validators.
Limit to keep in view
The environments imitate target websites; client-heavy behavior can be approximated or replaced with static surrogates, so task coverage bounds the fidelity claims.
arXiv:2611.06254
HarnessFlow: MCTS-Guided Harness Optimization for LLM Agents
HarnessFlow treats an LLM agent’s harness—the executable code that manages context, state, tools, and control flow—as an optimization target while keeping the model and task environment fixed. It organizes evaluated harnesses in a refinement tree and uses an MCTS-inspired policy to choose which candidate to refine; a coding agent makes the code changes using archived implementations, evaluation feedback, and reusable components. With 15 refinement attempts per benchmark, the authors report the highest mean test score among compared methods on all six benchmarks, averaging 69.48%. The approach improves the average over the initial harness by 10.89 percentage points and over Meta-Harness by 3.95 points.
How it works
For a fixed model and environment, the paper defines harness performance as expected task reward and seeks a harness with the highest validation score. Candidate evaluation averages repeated executions; execution traces supplement the scalar score by helping diagnose failures. Search starts from an initial harness, spends a fixed number of refinement attempts, and selects the best observed validation candidate for evaluation on a disjoint test set.
What the paper evaluates
The evaluation covers six benchmarks: TAU2-bench and AppWorld for agentic tasks; SWE-bench Verified and BigCodeBench-Hard for code; and WebShop and BrowseComp-Plus for web interaction. The paper uses task success rate for TAU2-bench and AppWorld, task score for WebShop, resolution rate for SWE-bench Verified, Pass@1 for BigCodeBench-Hard, and answer accuracy for BrowseComp-Plus. Each benchmark has a fixed-seed validation/test split, and the test set is not exposed during search. Methods use GPT-5.4-mini at medium reasoning effort in the main setup, five runs, and 15 candidate attempts.
Limit to keep in view
The authors state that the validation-score formulation requires tasks with an automatic grader; applicability to tasks without one is not established.
Read the full reviews
- PSGAgent: Enhancing MLLMs' Dynamic Visual Perception for Panoptic Scene Graph Generation via Agentic Reasoning · arXiv:2611.07408
- MRC-MMAgent: Reinforcement Learning for Multi-Role Collaboration in Multimodal Tool-Using Agents · arXiv:2610.39324
- WebHarbor: Docking Real Websites for Evolving GUI Agent Environments · arXiv:2611.03634
- HarnessFlow: MCTS-Guided Harness Optimization for LLM Agents · arXiv:2611.06254
Source note: findings and limitations above come from the linked full paper reviews. Follow each original arXiv paper for the authors' complete methods and results.