Princeton's Mengdi Wang: LLMs haven't made a real scientific discovery yet

Share
Princeton's Mengdi Wang: LLMs haven't made a real scientific discovery yet

The models that cracked competition math still can't find a new particle — and Princeton's Mengdi Wang thinks she knows why.


Large models learn "the most likely" answer, and science lives in the long tail. That was the core of a keynote from Mengdi Wang, director of Princeton's Center for AI Innovation, at the Inclusion·Bund Conference in Shanghai on September 10: the training method behind today's LLMs has conquered math and coding but, in her words, still hasn't produced a genuinely new discovery in the fundamental sciences. Her reasoning is structural, not sentimental. Training optimizes toward the maximum of a probability distribution, so models overestimate common cases and underweight rare ones — but in physics, chemistry and biology, the frontier findings are exactly the rare ones, like spotting a star that was never in the training data.

The contrast with math and code comes down to verifiers. A compiler plus unit tests checks every program an agent writes, and formal proof systems like Lean check every step of a proof, so models can try, fail, and iterate against ground truth. Experimental science has no equivalent validator: results depend on personal judgment, specific instruments, and multi-team coordination — one survey in a Nature-family journal found roughly 70% of experiments fail to replicate, and about half the time even the original scientist can't reproduce their own result. Without a verifier, Wang argues, a model just keeps overfitting the known distribution instead of reaching past it.

She also pushed the argument somewhere darker: AI may be slowing science down, not speeding it up. Papers have grown exponentially since LLMs arrived, but the underlying observations of the physical world haven't increased — so the effective signal per paper is falling, and low signal-to-noise makes the real findings harder for anyone to find. Her fix is infrastructure, not model scale: her team is wrapping an automated graphene lab into callable APIs and building LabOS, an "operating system" for research labs where multimodal AI connects models, robots and real instruments, so every experiment is recorded, verifiable and replicable. The bet is that likelihood becomes possibility only when the physical world closes the loop.

We covered the flip side of this earlier this week — OpenAI's Astra cuts the bounded prime gap record to 186 — and the two stories bracket the same gap: verified, formal domains are where AI shines; the unverified physical world is where it stalls.

What to watch: whether verifier-backed setups like LabOS turn "AI for science" from paper generator into discovery engine — the first lab to make its experiments API-checkable wins the argument.

If AI slows science down by drowning signal in paper volume, what should a lab actually do differently? Write to us via our contact page.

Get SiliconNoon in your inbox

One email a day with the AI stories that matter — what happened, the numbers, and why it matters. No spam, unsubscribe in one click.

Read more

GigaChat 3.5 Reasoning lands under MIT — and beats DeepSeek on math

GigaChat 3.5 Reasoning lands under MIT — and beats DeepSeek on math

Sber's GigaChat team has open-weighted a 432-billion-parameter reasoning model under an MIT licence, and the numbers put it within touching distance of DeepSeek V4 Flash. GigaChat 3.5 Reasoning runs 28 billion active parameters, trains natively in FP8, and mixes multi-head latent attention with GatedDeltaNet linear-attention layers — a hybrid aimed squarely at long-context cost — with a 262K-token window and three speculative-decoding heads bolted on.

Envision's Tianji is now the best weather model nobody in the West talks about

Envision's Tianji is now the best weather model nobody in the West talks about

Three stories from Shanghai's Bund Summit and the Chinese feeds this morning: an energy company's weather model just took the crown on the field's most-cited benchmark, Alipay's assistant started booking real travel, and a Nobel laureate put a number on what AI is worth. Envision's Tianji weather model entered Google's WeatherBench 2 for the first time this week and landed in the top tier alongside Google's own WeatherNext 2 — beating it on wind speed and precipitation, the two variables the…

Fable 5.1 leads Real-SWE, and still fails 6 of 10 tasks

Fable 5.1 leads Real-SWE, and still fails 6 of 10 tasks

Two benchmark results landed the same day, and both point at the same gap: agents score well on code they have seen, and badly on code they have not. The best coding agent in the world resolved 38.8% of real enterprise engineering tasks — and no agent solved every task in the set. Specific Labs, a Y Combinator-backed evaluation shop, released Real-SWE this week, a benchmark whose defining property is that its code has never touched the public internet.

A small model that knows when to hand off cuts inference cost 96%

A small model that knows when to hand off cuts inference cost 96%

The cheapest way to get frontier-level math reasoning just got a working open-source demo: teach a 4B model to admit when it's stuck. PyroMind Dynamics says its PyroDash method lets a Qwen3.5-4B model solve hard reasoning benchmarks at near-frontier accuracy while paying roughly 4% of what a big model alone would cost. The trick is a control token. The small model streams an answer normally, and the moment it decides it can't finish, it emits <|llm_offload|>.