AmapAI

💻Coding the World into Amap🌍

[AmapAI @Alibaba_AMAP]

WSDM 2027[4/4]|GeoGR is an industrial-scale generative POI recommendation framework for navigation-based LBS, which builds geo-aware Semantic IDs through spatio-temporal collaborative contrastive learning, RQ-Kmeans, and EM-style refinement, then aligns them with an LLM through CPT and SFT, and has delivered significant online gains on the AMAP platform.

https://t.co/qjyI3OTy2S
#Recommender #Recommendation #Semantic #WSDM

[AmapAI @Alibaba_AMAP]

WSDM 2027[1/4]|This work establishes when multiple behavioral and textual proxies can support reliable explanations of user preferences, and when the available evidence calls for abstention.

https://t.co/NZVXZkOjQN
#ExplainableAI #UserPreferences #RobustAttribution #WSDM

[AmapAI @Alibaba_AMAP]

WSDM 2027[3/4]|HeMix is a scalable industrial ranking model that combines query-mixed interest extraction across long-term and real-time behavior sequences with a linear-complexity HeteroMixer block for heterogeneous feature interaction, scales smoothly from ~100M to ~1.5B parameters, and has delivered +0.88% GMV and +2.74% PV_CTR on the AMAP App.

https://t.co/cKnjdWJXbe
#Heterogeneous #Recommenders #WSDM

[AmapAI @Alibaba_AMAP]

NeurIPS'26[16/16] | VeriTrip
✅ Evidence-grounded travel planning over unstructured web data
✅ Cell-wise factual verification for reliable agent evaluation
✅ Revealing the trade-off between retrieval and instruction following
📝 https://t.co/HipNYDGLej
#AI #MultimodalAgents

[AmapAI @Alibaba_AMAP]

NeurIPS'26 [15/16]| ACTGUIDE-RL
✅ learning from action guidance to help LLM agents overcome exploration barriers and internalize the gains through RL.
✅ No SFT cold start needed.
✅ +10.7pp on GAIA and +19pp on XBench with Qwen3-4B comparing to Zero RL.
#AgenticRL #LLM

[AmapAI @Alibaba_AMAP]

NeurIPS'26[14/16] | Plan4D
PLAN4D: Generative Plannable 4D Worlds
Action-conditioned generation of future videos and dense 4D worlds.
✅ Joint RGB, depth, camera & scene-flow prediction
✅ Stitches video generation with pretrained 3D geometry
#NeurIPS2026 #WorldModels

[AmapAI @Alibaba_AMAP]

NeurIPS'26[1/16] | OmniVerifier-C1
Better rewards for image generation: decompose, verify, score.
✅ Graded rewards capture partial correctness
✅ Reference-guided verification for structured visual tasks
📈 FLUX GenEval: 0.68 → 0.79
#NeurIPS2026 #MultimodalAI

[AmapAI @Alibaba_AMAP]

NeurIPS'26 [13/16]| OneHRL
Hybrid Reinforcement Learning for One-step Degraded Infrared and Visible Image Fusion
One-step fusion meets hybrid reinforcement learning for robust perception under composite degradations.
📄 Paper: https://t.co/aJToiTGLoC
#NeurIPS2026 #ImageFusion

[AmapAI @Alibaba_AMAP]

NeurIPS'26 [12/16]| DSI
Breaking the Static: Dynamic Text Conditioning for Diverse Image Generation
✅ Static prompts restrict early structural exploration
✅ DSI aligns text guidance with denoising stages
✅ Diversity improves while alignment and aesthetics largely hold
#Image

[AmapAI @Alibaba_AMAP]

NeurIPS'26[11/16] | Stable3R
Stable3R: Streaming 3D Reconstruction with Stable Geometric References
https://t.co/urCrl3y7X9
#3DReconstruction #SpatialAI #NeurIPS2026

[AmapAI @Alibaba_AMAP]

NeurIPS'26 [10/16]| ImagineAgent
Teaching agents to “imagine” new viewpoints for better open-vocabulary human–object interaction understanding.
Paper: https://t.co/dRF8YzitR4
Tool-augmented RL × Generative world modeling
#NeurIPS2026 #MultimodalAI #HumanObjectInteraction

[AmapAI @Alibaba_AMAP]

NeurIPS'26[9/16] | PRPO
Reinforce visually grounded, noise-stable tokens for better multimodal reasoning.
https://t.co/vJvVhVSjHe
#NeurIPS2026 #MultimodalAI #RL

[AmapAI @Alibaba_AMAP]

NeurIPS'26 [8/16]| Urbex
Ground city-scale 3D scenes via landmarks, zoom and oblique views.
https://t.co/UeHfZMJH6V
#SpatialAI #3DGrounding #NeurIPS2026

[AmapAI @Alibaba_AMAP]

NeurIPS'26[7/16] | DriveStreamBench
User-conditioned driving alerts, evaluated for timing and silence.
https://t.co/kHy7q5WSDL
#StreamingVideo #VLM #NeurIPS2026

[AmapAI @Alibaba_AMAP]

NeurIPS'26[6/16] | ALAM
✅ Algebraically consistent latent actions
✅ MetaWorld: 47.9%→85.0%
✅ LIBERO: 94.1%→98.1%
📝 https://t.co/M2pJMbkaZ2
#Robotics #VLA #EmbodiedAI #LatentAction

[AmapAI @Alibaba_AMAP]

NeurIPS'26[5/16] | MPG
Aug-free masked pretraining for pixel-space diffusion.
✅ +15.1% avg FID gain on 7 transfer datasets
✅ NC-CKNNA noise-trajectory metric
#NeurIPS2026 #Diffusion #ImageGen

[AmapAI @Alibaba_AMAP]

NeurIPS'26 [4/16]|SaaSBench
✅ From-scratch enterprise SaaS benchmark
✅ Dependency-aware hybrid evaluation
✅ Most failures occur before deep business logic
📝 https://t.co/2TobPBU9Ga
#CodingAgents #SaaS #LLM #Agentic

[AmapAI @Alibaba_AMAP]

NeurIPS'26 [3/16]| TransitLM
✅ End-to-end transit routes, without maps or routing engines
✅ GPS-to-station grounding through learned spatial knowledge
✅ One model for optimal, preferred, and multiple routes
📝 https://t.co/rQIR8l1Fb8
#LLM #PublicTransit #AI短剧

[AmapAI @Alibaba_AMAP]

NeurIPS'26[2/16] | DiGPro
Less forgetting via adaptive gradient projection.
No past-task data. MLLM-DCL: 65.34% avg accuracy, 2.27% avg forgetting.
#ContinualLearning #MultimodalAI #NeurIPS2026