BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//pretalx//pretalx.adass.org//adass2026//speaker//XQ8SAB
BEGIN:VTIMEZONE
TZID:AWST
BEGIN:STANDARD
DTSTART:20000101T000000
RRULE:FREQ=YEARLY;BYMONTH=1;UNTIL=20051231T160000Z
TZNAME:AWST
TZOFFSETFROM:+0800
TZOFFSETTO:+0800
END:STANDARD
BEGIN:STANDARD
DTSTART:20070325T040000
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=3
TZNAME:AWST
TZOFFSETFROM:+0900
TZOFFSETTO:+0800
END:STANDARD
BEGIN:DAYLIGHT
DTSTART:20071028T030000
RRULE:FREQ=YEARLY;BYDAY=4SU;BYMONTH=10;UNTIL=20081025T190000Z
TZNAME:AWDT
TZOFFSETFROM:+0800
TZOFFSETTO:+0900
END:DAYLIGHT
END:VTIMEZONE
BEGIN:VEVENT
UID:pretalx-adass2026-ZS9EYM@pretalx.adass.org
DTSTART;TZID=AWST:20261102T160000
DTEND;TZID=AWST:20261102T161500
DESCRIPTION:Radio interferometric pipelines encode expert judgment as fixed
  heuristics: reasonable defaults for a "typical" dataset. When a dataset d
 oesn't fit the mold\, an expert inspects the diagnostics and steps in by h
 and. That doesn't scale as surveys push toward higher data volumes and les
 s human oversight. LLMs can close that gap. They read the same diagnostics
  an expert would and reason about them per dataset. But letting a model bo
 th reason and act is risky. LLMs hallucinate\, drift\, and skip steps sile
 ntly. We built an architecture that strictly separates measurement from re
 asoning. The model gets room to adapt\, but it never touches the data path
  directly.\n\nAn MCP layer wraps the reduction software (CASA\, here) and 
 exposes each Measurement Set operation (metadata queries\, instrument geom
 etry\, calibration\, imaging) as an independent tool. Every tool returns s
 tructured data with explicit completeness and provenance. None of them int
 erpret their own output or call each other. Reasoning lives outside the to
 ols\, in skills: version-controlled\, plain-text documents that encode int
 erferometric expertise. They're fed into the LLM's context stage by stage\
 , so it can reason about what the tools hand back.\n\nAn orchestration lay
 er walks the model through a sequence of stages that looks like a processi
 ng pipeline\, except the parameters aren't fixed. Skill-based reasoning le
 ts the model make an informed\, per-dataset call instead of falling back o
 n defaults. Every stage's state\, its parameter choices\, and the sequenci
 ng are written out as documentation and a reproducible script. Since reaso
 ning lives outside the MCP layer\, the orchestrator doesn't care which mod
 el is driving it. Cloud (Claude\, Codex) or local and open (Gemma\, Qwen):
  both work. A model equipped with real domain expertise\, via skills\, bea
 ts a one-size-fits-all pipeline\, dataset by dataset.\n\nWe demonstrate en
 d-to-end inspection and calibration on VLA\, GMRT\, and ALMA data. We disc
 uss extending the approach to other instruments\, facilities\, and pipelin
 es.
DTSTAMP:20261116T152700Z
LOCATION:Banquet Hall
SUMMARY:LLM-Orchestrated Radio Interferometric Data Reduction - Srikrishna 
 Sekhar\, Preshanth Jagannathan
URL:https://pretalx.adass.org/adass2026/talk/ZS9EYM/
END:VEVENT
BEGIN:VEVENT
UID:pretalx-adass2026-VTTC3P@pretalx.adass.org
DTSTART;TZID=AWST:20261104T154800
DTEND;TZID=AWST:20261104T160000
DESCRIPTION:Faraday rotation along a line of sight is recovered either by F
 araday synthesis\, which transforms the Stokes Q and U spectra into Farada
 y depth space\, or by QU-fitting\, which fits parametric depolarization mo
 dels directly to Q(λ²) and U(λ²). QU-fitting returns calibrated uncert
 ainties and handles model comparison naturally\, but it costs one full Bay
 esian inference per sightline\, typically 10⁴ to 10⁶ forward-model eva
 luations under a sampler. At survey scale this is prohibitive.\n\nVROOM-SB
 I removes the sampler. Simulation-based inference never evaluates a likeli
 hood. We draw parameters from the prior\, push them through the same forwa
 rd model QU-fitting already uses (depolarization model\, frequency coverag
 e\, noise realization)\, and keep the resulting parameter and spectrum pai
 rs. A neural density estimator learns the mapping from the simulations the
 mselves\, so anything the simulator can generate can be inferred\, includi
 ng instrumental effects that are awkward to write into an analytic likelih
 ood.\n\nThe specific method is neural posterior estimation. A conditional 
 normalizing flow q(θ | Q\, U) is trained on the simulated pairs by maximi
 zing the log-density of the true parameters given their spectra\, so the n
 etwork targets the posterior directly rather than a likelihood or a summar
 y-statistic distance. Training is paid once per depolarization model and f
 requency setup. Inference on a new sightline is then a single forward pass
  returning a full posterior in milliseconds. This amortization is the enti
 re source of the 500x speedup over traditional QU-fitting. Our cost scales
  with the number of models trained\, not the number of pixels.\n\nRecovere
 d posteriors agree with Faraday synthesis and classical QU-fitting carried
  out using RM-Tools\, with coverage validated by Tarp testing and simulati
 on based calibration\, at roughly 500× lower per-sightline cost. We have 
 released the code on github (https://github.com/skunkworks-ra/vroom-sbi) a
 nd trained models via HuggingFace (https://huggingface.co/skunkworks-ra/vr
 oomsbi) currently for L-Band at the VLA.
DTSTAMP:20261116T152700Z
LOCATION:Seminar Room 2
SUMMARY:VROOM-SBI: Amortized Bayesian QU-Fitting for Survey-Scale Faraday A
 nalysis - Preshanth Jagannathan\, Arpan
URL:https://pretalx.adass.org/adass2026/talk/VTTC3P/
END:VEVENT
END:VCALENDAR
