ravikant.dev · Papers

Frontier LLM Papers — the bold ideas

Deep, honest explainers of recent papers I find genuinely interesting. The lens is narrow on purpose: giving LLMs intrinsic motivation (the closest thing to emotions / internal drives) and RL loops built on simple reward signals instead of elaborate RLVR machinery. Efficiency tweaks and incremental gains are deliberately excluded.

What this is

Each card links to a full walkthrough: the bold idea, exactly what they did (the real method, with the reward definition), the results with real numbers, and my honest take on whether it holds up. Curated by taste, not comprehensive — and shaped by your feedback at the bottom of every page.

Overviews of the intrinsic-reward field

🔥 The approaches & the math that separates them ↗ a synthesis by method (surprise, novelty, information gain, learning progress, empowerment, homeostasis) — the actual reward formula for each, explained simply, with the papers that use it. Start here to understand the subject.

📚 The ranked field map — 84 papers ↗ the whole literature as one top-to-bottom reading list.

🗺️ Prior art & landmines — 204 papers ↗ the continual-learning side of the map: continual pretraining, learnability-based data selection, self-improving agents, and every documented failure mode — ranked into 4 tiers with a "why it matters" note per paper.

Up next to read

Also on the shortlist (not written up yet)

The landscape is bigger than 10. These were strong candidates I cut for space or overlap — tell me (below) if you want any promoted:

R-Zero — co-evolving Challenger/Solver, zero data (close cousin of Absolute Zero).
HEART — emotionally-driven test-time scaling: emotional cues steer self-correction.
i-MENTOR / Navigate the Unknown — dense intrinsic exploration rewards for reasoning.
ONI — learn an intrinsic reward online from LLM feedback for RL agents.
RLP — reward the information-gain of the chain-of-thought at pretraining (successor to RPT).
SPICE — self-play in corpus environments improves reasoning.

Shape the next round

This whole collection is a feedback loop. Tell me what to cover next and which themes to lean into. Anonymous.

Which themes do you want more of?

A paper (or direction) I should explain next

Anything else — on the picks or how I explain them

Thanks — logged. I'll fold this into the next round.