BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//pretalx//pretalx.iacapconf.org//iacap-2026//talk//T89WHU
BEGIN:VTIMEZONE
TZID:US/Central
BEGIN:DAYLIGHT
DTSTART:20250716T000000
TZNAME:CDT
TZOFFSETFROM:-0500
TZOFFSETTO:-0500
END:DAYLIGHT
BEGIN:STANDARD
DTSTART:20251102T020000
RDATE:20261101T020000
TZNAME:CST
TZOFFSETFROM:-0500
TZOFFSETTO:-0600
END:STANDARD
BEGIN:DAYLIGHT
DTSTART:20260308T030000
RDATE:20270314T030000
TZNAME:CDT
TZOFFSETFROM:-0600
TZOFFSETTO:-0500
END:DAYLIGHT
END:VTIMEZONE
BEGIN:VEVENT
SUMMARY:Hallucination as a Property of Processing Mode - Lev Sukherman
DTSTART;TZID=US/Central:20260716T154500
DTEND;TZID=US/Central:20260716T161500
DTSTAMP:20260726T081834Z
UID:pretalx-iacap-2026-T89WHU@pretalx.iacapconf.org
DESCRIPTION:Large language model (LLM) hallucinations are typically treate
 d as architectural defects requiring technical fixes. We argue instead tha
 t hallucinations are a systemic property of the _syntactic processing mode
 _\, in which a system manipulates symbols without access to their semantic
  content. Our analysis focuses on base autoregressive architectures\; hybr
 id systems incorporating explicit semantic representations or external ver
 ification may partially mitigate the effects we describe\, but do not elim
 inate the underlying syntactic processing mode in the core generation mech
 anism. This perspective connects Searle's Chinese Room argument with Harna
 d's symbol grounding problem and empirical findings from cognitive psychol
 ogy. We demonstrate that humans operating in syntactic mode produce errors
  that share key properties with LLM hallucinations: high confidence\, surf
 ace plausibility\, and reliance on pattern-matching over verification. In 
 the Wason Selection Task\, approximately 10% of subjects arrive at the log
 ically correct answer\, yet the majority report high confidence in their i
 ncorrect responses\, much like LLMs producing false statements with high p
 robability scores. Cheng and Holyoak show that accuracy increases progress
 ively as semantic grounding moves from full abstraction to goal-directed c
 oncretization. Cosmides and Tooby extend this gradient to its furthest poi
 nt: when the same logical task is recast as a social contract\, human accu
 racy jumps to approximately 75%\, revealing a capacity for mode-switching 
 that LLMs lack. This reframes the hallucination problem: both humans and L
 LMs hallucinate when processing symbols without semantic grounding\, and i
 t is the processing mode\, rather than the agent\, that is the source of e
 rror.
LOCATION:Executive Conference Room
URL:https://pretalx.iacapconf.org/iacap-2026/talk/T89WHU/
END:VEVENT
END:VCALENDAR