HKU Seminar Argues LLMs' Math Success Is Fragile: Inductive Engines Can't Reliably Do Deduction
YiMaTweets · x · 2026-09-07
HKU IDS Scholar Seminar #28, "LLM Math: Prospects and Challenges" (Sep 25, 2026), will be given by Prof Boris Babic of HKU's Philosophy Department.
- Core thesis: training LLMs to excel at formal reasoning requires a predominantly inductive inference engine to produce deductively valid ideas — a structural mismatch
- This yields four persistent failures: reasoning fragility, parsing instability, poor numerology, and external tool dependence
- Babic argues these failures are predictable and enduring, and proposes "internalist guarantees" of LLM algorithmic performance via direct learning, manual coding, or outsourcing
- Based on a PNAS publication by Prof Qingpeng Zhang's group, first author Dr Fei Jing
More from AGI Musings
- Cloudflare exec: agentic traffic will be 1,000x human traffic in five years — 4KTV · 2026-09-07
- Reddit Mocks 'Astra Is AGI Because It Can Use Blender' Hype — Possible_Door_9719 · 2026-09-07
- Mathematician clarifies: some AI-generated math results do have value — littmath · 2026-09-07
- Paras Chopra: Build AI as Human Exoskeletons, Not Full-Autonomy Replacements — paraschopra · 2026-09-07
- AI makes verifiable work matter: the easier to verify, the bigger the contribution — RexDouglass · 2026-09-07
- Toby Walsh's God AI: superintelligence may arrive within our lifetime — TobyWalsh · 2026-09-07