BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:Failure-Aware LLM Serving for Agentic Workloads
DTSTART:20260617T090000
DTEND:20260617T110000
DTSTAMP:20260925T174510Z
UID:8a93657bc7a693269a8744d70ccf93b4f6f282ae2945a511f2bef439
CATEGORIES:Conferences - Seminars
DESCRIPTION:Palak\nEDIC candidacy exam\nExam president: Prof. Anastasia Ai
 lamaki\nThesis advisor: Prof. Anne-Marie Kermarrec\nCo-examiner: Prof. Rob
 ert West\n\nAbstract\nLLM deployment has shifted from single-turn question
  answering to multi-step agentic tasks\, where a model reasons\, acts\, an
 d observes in a loop until a complex goal is achieved. Serving these workl
 oads at scale introduces challenges that existing inference infrastructure
  was not designed to address. This report examines three works that collec
 tively define the landscape of this problem and motivate a thesis directio
 n centered on failure-aware serving for agentic workloads. ReAct establish
 es the execution model underlying production agentic systems\, showing tha
 t interleaving reasoning and acting in a loop produces observable executio
 n trajectories that reflect the agent's reasoning process at each step. Ag
 entix demonstrates that scheduling at the agentic task level rather than t
 he individual LLM call level yields substantial throughput improvements\, 
 but assumes all running tasks are making progress toward a correct complet
 ion. MAST shows this assumption does not hold: failure rates are strikingl
 y high even across state-of-the-art systems\, and many failure modes leave
  structurally observable patterns in execution trajectories. We present Or
 bit\, a trajectory-aware scheduler that acts on this observation by dynami
 cally deprioritizing tasks whose trajectories exhibit failure signals. The
  broader thesis outlines a principled framework for failure-aware serving 
 for agentic workloads\, developing the capabilities required to detect fai
 lure signals from partial trajectories online\, calibrate them to specific
  deployments\, and act on them effectively across diverse agentic workload
 s.\n\nSelected papers\n\n	ReAct (https://arxiv.org/pdf/2210.03629)\n	Agent
 ix (https://www.usenix.org/system/files/nsdi26-luo.pdf)\n	Why do multi-age
 nt systems fail? (https://arxiv.org/pdf/2503.13657)\n
LOCATION:BC 133 https://plan.epfl.ch/?room==BC%20133
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR