Sharad Agarwal

Sharad Agarwal, Pooria Namyar, Alec Wolman et al.

Agentic AI systems that invoke external tools are powerful but costly, leading developers to default to large models and overspend inference budgets. Model routing can mitigate this, but existing routers are designed for chat completion rather than tool use. We present Switchcraft, the first (to the best of our knowledge) model router optimized for agentic tool calling. Switchcraft operates inline, selecting the lowest-cost model subject to correctness. We construct an evaluation framework on five function-calling benchmarks and train a DistilBERT-based classifier, deployed under a latency budget. Switchcraft achieves 82.9% accuracy -- matching or exceeding the best individual model -- while reducing inference cost by 84%, saving over $3,600 per million queries. We find that larger models do not consistently outperform smaller ones on tool-use tasks, and that nominally cheaper models can incur higher total cost due to token-intensive reasoning. Our work enables cost-aware agentic AI deployment without sacrificing correctness.

Yuxin, Liu, Yoshimichi Nakatsuka et al.

Demonstrating the veracity of videos is a longstanding problem that has recently become more urgent and acute. It is extremely hard to accurately detect manipulated videos using content analysis, especially in the face of subtle, yet effective, manipulations, such as frame rate changes or skin tone adjustments. One prominent alternative to content analysis is to securely embed provenance information into videos. However, prior approaches have poor performance and/or granularity that is too coarse. To this end, we construct Vronicle -- a video provenance system that offers fine-grained provenance information and substantially better performance. It allows a video consumer to authenticate the camera that originated the video and the exact sequence of video filters that were subsequently applied to it. Vronicle exploits the increasing popularity and availability of Trusted Execution Environments (TEEs) on many types of computing platforms. One contribution of Vronicle is the design of provenance information that allows the consumer to verify various aspects of the video, thereby defeating numerous fake-video creation methods. Vronicle's adversarial model allows for a powerful adversary that can manipulate the video (e.g., in transit) and the software state outside the TEE. Another contribution is the use of fixed-function Intel SGX enclaves to post-process videos. This design facilitates verification of provenance information. We present a prototype implementation of Vronicle (to be open sourced), which relies on current technologies, making it readily deployable. Our evaluation demonstrates that Vronicle's performance is well-suited for offline use-cases.

Jessica Tang, Shraddha Barke, Sharad Agarwal

When a language model produces a response in a multi-turn conversation, which tokens from prior turns shaped that answer, and how did those dependencies propagate across prior turns? Existing context attribution methods process the full context in a single pass, recovering surface-level dependencies but missing the layered, non-linear structure of real-world dialogues and multi-step reasoning tasks. We introduce multi-turn context attribution (MTCA): given a target span in a model response, the task of tracing attribution backward across turns to identify not only which prior turns were directly relevant, but also how those turns themselves depended on earlier context. We propose Tokengeist, an attribution-method-agnostic and scalable framework that recovers full dependency paths by casting attribution as a recursive traversal of a directed acyclic graph (DAG) over conversation turns. We will release MTCABench, a benchmark of 3,845 target spans across 665 multi-turn conversations, annotated with gold provenance graphs reaching depths of up to 14, across four dependency types. Across four open-weight models, flat attribution methods fail to recover multi-hop dependencies, achieving under 20% source recall, while Tokengeist reaches 90%. Our results reveal systematic failure modes of single-pass attribution -- which we term provenance collapse -- and motivate attribution methods that reason recursively across turns.