272

Papers

10/10

Roadblocks Active

0

Connections

MolmoWeb trains 4B and 8B vision-language models to control web browsers using only screenshots—no HTML or accessibility tree—by learning from a mix of 100K+ synthetic trajectories and 30K+ human demonstrations. The 8B model outperforms GPT-4o-based agents on WebVoyager, and parallel best-of-4 sampling pushes pass rates from 78% to 95%. This matters because it demonstrates that open, modest-scale models can match frontier closed systems for web automation when training data quality and diversity are prioritized.

██████████ 0.9 agent-tool-use Preprint

ReadSaveConnections

This benchmark tests whether current reward models—generative, discriminative, and LLM-as-judge—can evaluate complete multi-step agent trajectories and finds all three families fail substantially, especially on long-horizon tasks and complex tool use. This is the missing infrastructure problem for aligning agents: without reliable trajectory-level signals, RLHF-style training cannot safely shape agent behavior. The result implies that scaling agent capabilities without solving reward modeling first is building on an unstable foundation.

██████████ 0.9 agent-tool-use Preprint

ReadSaveConnections

Current multimodal agents call external tools even when their internal knowledge is sufficient, creating latency and reasoning noise. The paper introduces HDPO (Hierarchical Decoupled Policy Optimization), which trains accuracy and efficiency as separate RL objectives—efficiency rewards are only applied to rollouts that were already correct, preventing the agent from learning to skip tools by getting answers wrong. The resulting Metis-8B model and weights are publicly released, making this directly usable for practitioners.

██████████ 0.9 agent-tool-use Preprint

ReadSaveConnections

HiRO-Nav finds that only a small fraction of navigation steps are genuinely uncertain (high action entropy), and these are precisely the moments where activating slow chain-of-thought reasoning most improves task success. By using entropy as a gate to selectively trigger deep reasoning—and training via a two-stage RL pipeline with KL regularization to prevent forgetting—the system achieves better success rates than always-thinking or never-thinking baselines while using far fewer tokens per episode. This selective-reasoning idea is broadly applicable to any agentic VLM system where inference cost matters.

██████████ 0.9 embodied-ai Preprint

ReadSaveConnections

MedVR trains a 7B medical vision-language model without any human-labeled reasoning steps, using two mechanisms: entropy-guided visual regrounding (using the model's own prediction uncertainty to re-examine ambiguous image regions) and consensus-based credit assignment (aggregating bounding boxes across diverse rollouts to generate pseudo-labels). It achieves state-of-the-art results on six public medical VQA benchmarks including out-of-domain tests, and code is publicly released. This is notable because labeling intermediate reasoning steps in medical imaging is extremely expensive, making annotation-free RL approaches practically important.

██████████ 0.9 hallucination-grounding Preprint

ReadSaveConnections

When fine-tuning a single model across multiple task types (math, spatial reasoning, visual grounding), standard RL training is dominated by whichever task produces the largest advantage values—starving other tasks of gradient signal. G²RPO fixes this by using 1D optimal transport to force each task's advantage distribution to match a standard normal, ensuring equitable gradient contribution across domains. Evaluated on 18 benchmarks across 6 categories on a Qwen3-VL-8B base; the approach is task-agnostic and could apply to any multi-domain RL fine-tuning scenario.

██████████ 0.8 multimodal-understanding Preprint

ReadSaveConnections

The paper finds that over 47% of image editing instructions generated by strong baseline vision-language models contain critical errors—mainly wrong spatial orientation, viewpoint ambiguity, or missing attribute detail—which pollutes training data for downstream image editors. A two-stage pipeline (SFT on 100K filtered examples, then DPO on 10K human preference pairs) cuts critical errors from 47.75% to 23% and raises correctness from 42% to 66%. This highlights that data quality for visual instruction synthesis is a concrete, measurable bottleneck, not just an abstract concern.

██████████ 0.8 multimodal-understanding Preprint

ReadSaveConnections

Using a 3D Pokémon game environment where agents see only raw RGB pixels and must complete long-horizon tasks, the benchmark finds that physical deadlock recovery—getting unstuck from environmental traps—is a stronger predictor of task failure than high-level planning. A surprising metacognitive split emerges: weaker models enter deadlocks without realizing it, while stronger models detect the deadlock but still cannot escape. This suggests that physical world modeling and self-correction, not just reasoning ability, are critical next targets for VLM development.

██████████ 0.8 multimodal-understanding Preprint

ReadSaveConnections

Network traffic classification today is purely pattern-matching—it cannot explain why a packet sequence looks suspicious in human-readable terms. This paper builds BGTD, a benchmark pairing raw traffic bytes with structured semantic annotations generated via Claude Opus across six public repositories, and trains mmTraffic, a model that combines a traffic encoder with an LLM to produce auditable reasoning reports while maintaining competitive classification accuracy. The interpretability angle is practically significant for security operations where explainability is a regulatory requirement.

██████████ 0.8 hallucination-grounding Preprint

ReadSaveConnections

Direct LLM planning for IoT sensor scheduling fails because LLMs lack structured spatial and semantic representations of sensor networks. IoT-Brain introduces a Spatial Trajectory Graph (STG) as a neuro-symbolic intermediate, combined with a verify-before-commit discipline that checks proposed schedules against physical constraints before execution. The result is a 37.6% higher task success rate than the best search-based methods while using 6.6x fewer prompt tokens—concrete evidence that structured grounding outperforms raw LLM planning for constrained optimization problems.

██████████ 0.8 reasoning-reliability Preprint

ReadSaveConnections