ACM TechTalk: From Conventional LLMs to Reasoning Models to Agents
Logistics
- Event
- ACM TechTalk (ACM Learning Webinar) — From Conventional LLMs to Reasoning Models to Agents
- When
- Wednesday, 2026-08-19. Time not stated in the announcement — confirm on the Zoom registration page (ACM TechTalks usually run 12:00 ET).
- Where
- Virtual (Zoom Events)
- Speaker
- Sebastian Raschka
- Announcement
- on.acm.org t/3538 (posted 2026-07-23, edited same day)
- Registration
- Zoom Events
- Cost
- Free; recording normally posted to the ACM Learning Center afterwards
Abstract (as posted)
Many of us have been using and experimenting with LLMs for several years. In 2025 new techniques for training reasoning models gave LLMs a noticeable boost in problem-solving; since then newer generations have been substantially more useful for complex tasks such as math and coding. In software development the progression has been from code completion to planning, implementing, testing, and debugging entire features and codebases.
The talk covers (a) how reasoning models differ from conventional LLMs, with a focus on inference scaling and training techniques, and (b) the main components of AI agents, showing how reasoning models serve as the engine that lets agents plan, use tools, evaluate intermediate results, and recover from mistakes.
Speaker people
Writes the Ahead of AI newsletter and the (From Scratch) book series. The
talk maps almost exactly onto Build a Reasoning Model (From Scratch) —
evaluation, inference-time scaling, RL, distillation — so the book's repo is
the executable version of the slides.
- GitHub
- rasbt. Directly relevant: reasoning-from-scratch (the reasoning
book's code, also on PyPI as
reasoning-from-scratch) and LLMs-from-scratch. Unlike most talks in this series, this one is backed by open code. - Adjacent 2026 talk
- PyCon DE 2026 keynote — "LLMs in 2026: From Architecture to Production". Worth diffing against this abstract to see which parts are the standing deck and which are new.
References
Primary — speaker's own work
| Item | Form | Link |
|---|---|---|
| Build a Reasoning Model (From Scratch) | Manning book, 440pp | manning.com |
| ⤷ code repo | PyTorch, runs on consumer hardware | rasbt/reasoning-from-scratch |
| ⤷ chapter notebooks | rendered HTML | sebastianraschka.com/reasoning-from-scratch |
| Build a Large Language Model (From Scratch) | Manning book | rasbt/LLMs-from-scratch |
| Ahead of AI | newsletter | magazine.sebastianraschka.com |
The reasoning book starts from a pretrained open base model (Qwen3) and adds reasoning post-hoc in four moves: evaluation, inference-time scaling, reinforcement learning, distillation. That decomposition is the likely spine of the talk.
As listed in the ACM announcement
Member-gated (ACM Skills Bundle); the O'Reilly entries are the same two Manning books above.
- Build a Reasoning Model (From Scratch) — O'Reilly, ISBN 9781633434677
- Build a Large Language Model (From Scratch) — O'Reilly, ISBN 9781633437166
- GenAI & Azure OpenAI Service: Chat and Reasoning Models in Foundry Playgrounds — Percipio
- Introduction to DeepSeek-R1 and Reasoning Models — Percipio
- Agentic AI Mastery — Percipio
Related ACM TechTalks (same category)
Prior art in the ACM Learning Webinars category, useful as a baseline for how much the framing has moved:
- Inside an LLM Agent: A From-Scratch Walkthrough (2026-05-07)
- Building Reliable AI Agents and LLM apps Using LangChain and LangGraph (2025-04)
- Understanding the LLM Development Cycle: Building, Training, and Finetuning (2024-05) — also Raschka
- Research to Reality: Building Production-Ready LLM Apps Users Can Trust (2024-11)
Related notes crosslink
- ACM, same season
- ACM AI Leadership Summit 2026 (Aug 30–Sep 2, Atlanta) — the Sep 2 Agentic AI track is the governance-flavoured counterpart to this talk's capability-flavoured framing.
- Agents
- agentic-2026 · 2026-q2-skills (harness/context/sandbox lens) · 2026-agent-memory-systems.
- Evaluation
- terminal-bench · llm-evaluation · llm-model-comparison-2024.
- Architecture
- attention-is-all-you-need · 2025-formal-transformer-verification · fine-tuning-machine-learning-models.
- Same-week meetups on the agent-harness question
- AICamp Boston (2026-07-23) · Boston Generative AI — The Harness for AI Agents (2026-07-23).
Claims to test refutation
- A reasoning model is an evolution of a conventional LLM, not a distinct architecture :: a demonstrated architectural change (not just training or decoding) that is required for the reasoning gain.
- a task family where longer chains lower accuracy (overthinking / self-distraction) at fixed model.
- an agent whose task success is dominated by harness quality (tools, retries, context) and barely moves when the reasoning model is swapped — the "harness, not model" thesis from the 2026-07-23 Boston meetups.
- recovery that is actually implemented in the scaffold (retry loop, verifier, compensating action) rather than emitted by the model.
Open questions / Q&A
- Where is the boundary between trained-in reasoning (RL, distillation) and decoded-in reasoning (best-of-N, self-consistency, budget forcing)? Which one carries the 2025 gains?
- Reward for RL: verifiable rewards (math/code unit tests) generalise how far outside verifiable domains?
- On distillation from a stronger reasoner — what survives, and does it transfer the failure modes too?
- Agent evaluation: is intermediate-result evaluation a separate verifier model, or the same model self-grading? Correlated-failure risk if the latter.
- Cost: at what token budget does a reasoning model stop being cheaper than a bigger conventional model on the same task?
Actions
[ ]Confirm the start time + timezone on the Zoom registration page[ ]Register[ ]Skimrasbt/reasoning-from-scratchchs. on inference scaling + RL before the talk[ ]Diff this abstract against the PyCon DE 2026 keynote description[ ]After: check the ACM Learning Center for the recording + slides