OSGR
A two-stage graph reasoning framework for Earth observation question answering. It narrows the search space through offline symbolic concept derivation, then instantiates observation items through online relation-conditioned grounding, mitigating the grounding failures and search-space explosion that LLM agents encounter under complex spatiotemporal constraints.
Timeline (last 90 days)
The OSGR (Observation-Support Graph Reasoning) framework addresses the retrieval and deduction challenges in Earth observation question answering through a two-stage decomposition of offline symbolic concept derivation and online relation-conditioned anchoring.
In the offline stage, OSGR constructs an Observation-Support Graph, which encodes support relations between concepts rather than storing concrete observation items, and is used to derive valid support paths from goals to results at the concept level.
In the online stage, OSGR employs a Relation-Conditioned Grounding mechanism, combining a spatiotemporal compatibility kernel to assess the degree of match between candidate observation items and query constraints, instantiating abstract concept paths into sets of concrete observation items.
On the two benchmarks Earth-Bench and DisasterM3-Reason, OSGR with GPT-4o raised accuracy from 49.3% to 63.7%.
On the two benchmarks Earth-Bench and DisasterM3-Reason, OSGR with GPT-4o reduced the execution step ratio from 2.9 to 2.5.
OSGR's validation is currently concentrated on academic benchmarks, and its generalization capability in real business scenarios (such as disaster emergency response and agricultural monitoring) remains to be tested.
The construction of OSGR offline graphs relies on conceptual-level knowledge engineering, and how to automatically extend the support graph to cover a broader range of geographic query types is a direction worthy of attention in the future.
All Facts (7)
The OSGR (Observation-Support Graph Reasoning) framework addresses the retrieval and deduction challenges in Earth observation question answering through a two-stage decomposition of offline symbolic concept derivation and online relation-conditioned anchoring.
50%UnverifiedIn the offline stage, OSGR constructs an Observation-Support Graph, which encodes support relations between concepts rather than storing concrete observation items, and is used to derive valid support paths from goals to results at the concept level.
50%UnverifiedIn the online stage, OSGR employs a Relation-Conditioned Grounding mechanism, combining a spatiotemporal compatibility kernel to assess the degree of match between candidate observation items and query constraints, instantiating abstract concept paths into sets of concrete observation items.
50%UnverifiedOn the two benchmarks Earth-Bench and DisasterM3-Reason, OSGR with GPT-4o raised accuracy from 49.3% to 63.7%.
50%UnverifiedOn the two benchmarks Earth-Bench and DisasterM3-Reason, OSGR with GPT-4o reduced the execution step ratio from 2.9 to 2.5.
50%UnverifiedOSGR's validation is currently concentrated on academic benchmarks, and its generalization capability in real business scenarios (such as disaster emergency response and agricultural monitoring) remains to be tested.
50%UnverifiedThe construction of OSGR offline graphs relies on conceptual-level knowledge engineering, and how to automatically extend the support graph to cover a broader range of geographic query types is a direction worthy of attention in the future.
50%