Why aren't LLM's trained on their own Chain Of Thought?

3•simianwords•1h ago

I realise that without allowing reasoning tokens, a model performs very poorly. It can't perform simple arithmetic or simple logic and hallucinates a bit.

But by allowing it to think a bit and then answer, the result is much better and way more trustable.

This shows a clean RL environment.. or just a nice data-set. Where you prompt the model two times - one without allowing thinking and one with thinking. Penalise the result from non thinking if the result contradicts the answer obtained from thinking.

Comments

i7l•1h ago

For the same reason that anyone's reasoning process and answers to random exam questions are never used as textbooks: if the reasoning is not guaranteed to be right, why would you want to make that training material?

simianwords•1h ago

We can empirically figure out how often the reasoning model is correct. With a 95% empirical accuracy, it should still help the model directionally. No training data set needs to be 100% accurate. No?

Rig: Build modular LLM apps in Rust – 20 providers, one unified interface

Ask HN: Will AI be the end of new programming languages?

Women were never meant to give birth on their backs

Show HN: Orcastrate – Sync GitHub Actions workflows across repos via templates

Mourning for dinosaurs, 65M years too late

The AI Compute Race: Microsoft's Miss and Oracle's Opportunity (2025)

Zipf's Law

Loqi, a memory system that preserves context after LLM compaction

The Forgotten Ones: Actron AM1608 16-Bit CPU. – The CPU Shack Museum

Apple at 50: My journey to the Mac – anderegg.ca

Show HN: Genetic algorithm engine that evolves trading strategies

Musician says AI company is cloning her music, filing claims against her

Emilia Britannia (public domain freedom mascot)

Show HN: RPLY - one Inbox for iMessage, WhatsApp, Slack, and Gmail on macOS

Track what top investors own (13F) and why they own it (10K AI Analysis)

Framework? I sure hope it does

Artemis II Tracker – Live Mission Control

Days Since OpenClaw CVE

Show HN: MailMark – Cold email tool where you own your domain and mailboxes

Show HN: Arbory – Native iOS dashboard and widgets for Plausible Analytics

An elegant Pomodoro timer for your terminal

/Render – 3D Model Skill for Claude Code

Can We Measure Software Slop? An Experiment

Chinese Chip Firms Hit Record High Revenue Driven by the AI Boom and U.S. Curbs

Why Your Engineering Team Is Slow (It's the Codebase, Not the People)

Journalist detained for booing Trump at Kennedy Center Chicago performance

Securing AI infrastructure to prevent backdoors and sabotage

Code Is Worthless

YAML is (not) my preferred configuration format

Eggplant