BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//CAIL//Events//EN
CALSCALE:GREGORIAN
METHOD:PUBLISH
X-WR-CALNAME:CAIL Events
BEGIN:VEVENT
UID:2026-11-13-qi-lei@cail.columbia.edu
DTSTAMP:20261110T200339Z
DTSTART:20261113T160000Z
DTEND:20261113T170000Z
SUMMARY:ML Seminar: Qi Lei - Reasoning in Two Dimensions: Quantifying Chai
 n-of-Thought Faithfulness and Efficiency
LOCATION:SSW Room 903\, School of Social Work\, 1255 Amsterdam Ave\, New Y
 ork\, NY 10027
GEO:40.81022;-73.958347
DESCRIPTION:Register: https://www.eventbrite.com/e/mlai-seminar-qi-lei-new
 -york-university-tickets-1999106771167\n\nChain-of-thought (CoT) reasoning
  exposes intermediate steps\, but a correct answer does not establish that
  an LLM relies on the dependence expressed by those steps or reasons effic
 iently. We operationalize CoT faithfulness and efficiency within a common 
 framework. Specifically\, we represent a CoT as a semantic reasoning graph
  and decompose the model’s predictive dependence along two complementary
  axes: dependence explained by semantic support versus additional reasonin
 g history\, and dependence allocated to answer-reaching versus off-path re
 asoning. This decomposition yields global and step-level measures of unfai
 thfulness and inefficiency. Controlled interventions show that the propose
 d metrics respond to hidden dependence and induced off-path reasoning as i
 ntended. Across mathematical and multi-hop reasoning tasks\, answer-level 
 unfaithfulness is consistently associated with greater predictive uncertai
 nty\, whereas both unfaithfulness and inefficiency exhibit weak or inconsi
 stent associations with answer correctness. We further derive theoreticall
 y grounded\, graph-aware fine-tuning objectives that improve the targeted 
 reasoning properties without directly optimizing the computationally expen
 sive metrics or compromising answer accuracy. Results indicate that accura
 cy and brevity provide incomplete assessments of reasoning quality: reliab
 le CoT evaluation should also consider whether the LLM’s predictive depe
 ndence is faithfully represented and efficiently allocated.\n\nhttps://cai
 l.columbia.edu/events/2026-11-13-qi-lei
URL:https://cail.columbia.edu/events/2026-11-13-qi-lei
END:VEVENT
END:VCALENDAR