因果推理

Certainty vs. Probability

All men are mortal.       12 × 12 = 144
Socrates is a man.        Not approximately. Not probably.
∴ Socrates is mortal.     Exactly 144. Necessarily.

Deductive reasoning produces certainty. The conclusion is forced by the structure. One misstep and the logic collapses. There is no "probably correct."

LLM prediction produces probability:

Given the distribution I have sampled during training, this is the token most likely masked or to follow.

Even at 99.99% confidence, it remains a statistical guess. A system trained to optimize for plausibility cannot, by design, produce necessity.

The question Weizenbaum asked in 1966 remains unanswered: Is what we are seeing intelligence, or a reflection of our desire to see it?

---

The Question

Do LLMs actually understand or do they predict plausible-sounding tokens without understanding?

This project surveys 360 papers to find out - tracking who supports the thesis, who challenges it, and what the evidence actually says.

To bring the findings home: - Paper network: interactive graph of 360 papers and 1353 relationships, filterable by stance - Experiments: - Decoding ablation: reasoning paths exist in base models, hidden by greedy decoding; RL surfaces them - Steering ablation: safety alignment is a thin layer of refusal patterns that washes off under trivial perturbations - Attractor states: extended LLM-to-LLM conversation reveals training distribution patterns (distribution chaos) - LLM Made Less Black Box: four visual explainers (Data → Tokenization → Architecture → Training) demystifying the full pipeline ---