The Big Picture
Agents that actively ask witnesses and check camera connectivity plus transition times find people across multiple cameras far more quickly and with fewer dialog turns than plain language-only methods.
ON THIS PAGE
The Evidence
Giving an agent tools to query a witness and a graph of where and how long it takes to move between cameras lets it rule out impossible candidates quickly. A single temporal check often eliminates most wrong matches, and a planner that chooses the right questions reduces wasted turns. Plain language-only inference fails to converge on many tasks; tool-augmented, grounded interaction is substantially more effective. tool-augmented, grounded interaction.
Not sure where to start?Get personalized recommendations
Data Highlights
1ARGOS contains 2,691 tasks across 3 tracks (989 in Track 1, 550 in Track 2, 1,152 in Track 3) built from 16-camera environments.
2Witness answers are “I’m not sure” about 85% of the time, because only 3 of 21 attributes are visually observable (≈14.3% expected answer rate).
3In one example, a single temporal feasibility check removed 16 of 19 candidates (≈84% reduction), showing high information value of camera timing checks.
What This Means
Engineers building agentic search systems and surveillance teams who need faster, more reliable person-finding across camera networks should pay attention—tools that combine questioning with camera layout and timing dramatically improve efficiency. Researchers and evaluators designing agent benchmarks can use ARGOS as a stress test for spatial and temporal reasoning abilities. stress test for spatial and temporal reasoning abilities.
Key Figures
Fig 1: Figure 1 : Overview. Left: An ARGOS agent interacts with an ambiguous witness through multi-turn dialogue, combining appearance, spatial, and temporal queries. Right: The four-module agent architecture (Analyst → \to Planner → \to Interviewer → \to Interpreter) forms an observe-think-act loop over the evaluation environment.
Fig 2: Figure 2 : Left: STTG for a 16-camera factory environment. Nodes are grouped into zones by OVERLAP connectivity. Edge types: OVERLAP (blue), SOFT_ADJ (orange), TRAVEL (gray); labels show median transition time. Right: 3D camera layout with sample imagery.
Fig 3: Figure 3 : Cumulative success rate (SR t @ T T ) vs. turn budget for Track 2 (left) and Track 3 (right), GPT-4o. The ARGOS Agent (blue) resolves tasks substantially faster than w/o Strategy (orange). w/o Spatial/Temporal Tool (red) stays near zero.
Fig 4: Figure 4 : Track 3: with vs. without temporal tool. A single temporal check eliminates 16 of 19 candidates by verifying spatio-temporal transition feasibility against the STTG, demonstrating the tool’s information density.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreConsiderations
The witness in ARGOS is a deterministic simulator, so real human memory errors and inconsistent answers are not represented. Benchmark tasks come from two environments (factory and university), so performance may change on different layouts or outdoor settings. The underlying capability has clear privacy and surveillance risks; evaluations are currently limited to consented public datasets and careful governance is needed for practical use. privacy and surveillance risks.
Methodology & More
Active, tool-augmented agents that mix targeted questions with spatial and timing checks solve multi-camera person search far more efficiently than language-only approaches. ARGOS frames the problem as multi-turn dialogue with a witness plus access to a spatio-temporal graph that encodes camera adjacency and transition-time statistics. The agent pipeline computes which visual attributes will best narrow candidates, plans the next action (ask about appearance, ask about location, or run a timing check), executes that action via a tool, and applies filters based on the reply. spatio-temporal graph The benchmark includes 2,691 tasks across three progressively harder tracks: single-turn attribute parsing, zone-level spatial disambiguation, and time-separated sightings that require temporal feasibility checks. Key findings: witnesses often cannot answer many attribute questions (≈85% uncertain answers), so relying on appearance alone stalls. Temporal checks grounded in the camera transition graph can eliminate most candidates in a single step (example: 16 out of 19 removed), and an agent that strategically selects questions reaches correct identifications in far fewer turns than unguided approaches. ARGOS also supplies turn-weighted success as a metric that rewards correctness and conversational efficiency, exposing parsing accuracy and tool-use strategy as the main bottlenecks for future work. Evaluation
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
Authors include researchers with h-index around 10 (mid-range); lacking prominent institutional signals or top-tier venue keeps it moderate.