Research map / Reinforcement learning / RL from feedback

FalconPerception-HD: High Density Perception via Reinforcement Learning

NeurIPS 2026 · Reinforcement learning › RL from feedback · approach: RL fine-tuning of generative models

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Applies GRPO-based RL post-training to the autoregressive Falcon Perception model to align open-vocabulary segmentation with set-matching metrics, removing the need for NMS and center deduplication and improving dense-scene recall.

Problem

Autoregressive perception models trained with per-token maximum likelihood are misaligned with set-level precision/recall metrics, leading to duplicate masks, sequence collapse in dense scenes, and reliance on NMS-style heuristics.

Approach

Use GRPO with a count reward (FN + 0.3·FP after Hungarian IoU matching) that samples only the LM and center heads while holding the segmentation head fixed, plus Dr. GRPO length-unbiased aggregation, Clip-Cov entropy control, and a stop-gradient on existence tokens to preserve MCC.

Key results

• On PBench, Falcon Perception-HD averages 59.4 F1 vs 56.8 for the SFT base, with the dense split rising from 72.3 to 80.5.
• On SACO, MCC improves from 59.6 to 62.2 with existence-token detachment, while removing NMS and dedup costs only 0.4 pmF1 (vs 12.0 for the base model).
• On dense scenes, the gain grows with object count, from +1.5 F1 at 0-2 objects to +21.5 F1 at 300+ objects.

Surprising findings

• Without stop-gradient on existence tokens, RL collapses MCC from 59.6 to 38.8 while inflating pmF1 and macro-F1.
• Sampling the size head adds no gain; the size head's entropy drops 21% despite receiving no gradients, attributed to a cascade effect.

Limitations

• Relies on the base model's own predictions and GPT-5 judging for data curation, plus human annotation for dense scenes.
• Only RL on positive queries; existence calibration depends on a stop-gradient heuristic.
• Dataset release is promised but not yet available ('Datasets will be published').

Experimental setup

Datasets: PBench SACO COCO VisDrone LVIS Dense200

Benchmarks: PBench SACO COCO-dense LVIS-dense Dense200 VisDrone COCO

Metrics: macro F1 F1@0.5 pmF1 MCC mask redundancy rate pass@k

Baselines: Falcon Perception SAM 3 Qwen3-VL Moondream3 Rex-Omni LocateAnything

Builds on: Falcon Perception GRPO Dr. GRPO VeRL Clip-Cov Muon

Closest papers by meaning

• Unleashing Perception-Time Scaling to Multimodal Reasoning Models (ICLR 2026 · similarity 0.88)
• OddGridBench: Exposing the Lack of Fine-Grained Visual Discrepancy Sensitivity in Multimodal Large Language Models (CVPR 2026 · similarity 0.87)
• ViCrit: A Verifiable Reinforcement Learning Proxy Task for Visual Perception in VLMs (NeurIPS 2025 · similarity 0.87)
• NoisyRollout: Reinforcing Visual Reasoning with Data Augmentation (NeurIPS 2025 · similarity 0.87)
• Perception-R1: Pioneering Perception Policy with Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
• Dr. Seg: Revisiting GRPO Training for Visual Large Language Models through Perception-Oriented Design (CVPR 2026 · similarity 0.86)