1The New Turing Test

At the AI Impact Summit in New Delhi, February 2026, Demis Hassabis proposed a test: train an AI system only on knowledge available up to 1911, and see if it can independently derive general relativity, just as what Albert Einstein did in 1915. He said it would be AGI if it can[1].

The idea has since been called the “Einstein test”, on which Nature ran a feature[2], and other researchers have turned it into a more formal evaluation framework or asked which skill AI is actually missing[3, 4]. This test asks the model to think of something nobody has thought of yet, highlighting the agent's ability to make scientific breakthrough, instead of simply pushing science forward within the existing framework. 70 years ago Turing used conversation to tell machines from people[5], while language models nowadays still struggles to fully imitate humans, whether models can pass this test is less cared about, as more hold the belief that super intelligence is not necessarily human like. We focus more and more on the utilitarianism side: whether this models can really change the world, hence the massive investment in coding agent, real-world interface, AI for sicence and also, corresponding benchmarks or tests like the Einstein test.

That said, this test is very hard to actually run. The main obstacle is models' prior knowledge. A modern language models already have read almost everything ever written about modern physics (and hopefully, in 100+ languages :/), so it is impossible to erase this knowledge from their wights and require them to "rediscover" the theory again.

A natural strategy is to train “vintage” models from scratch on old text only. These works include: talkie, a 13-billion-parameter model trained only on English text from before 1931 (you can chat with it online)[6]; the Ranke-4B family family, with cutoffs in various years from 1913 to 1946[7]; TimeCapsuleLLM, trained on London texts from 1800–1875[8]; and GPT-1900, built specifically for the Einstein test (called the Machina Mirabilis project in Nature’s report). Reportedly GPT-1900 once produced a passage resembling the 1905 light-quantum idea, but overall it was far from an “AI Einstein”[2].

Vintage models allows to investigate what does a model that has only read old books do, but unfortunately they cannot answer the question we actually care about: do today’s strongest models have what it takes to discover new physics? A 13B vintage model (even GPT-3 released in 2020 has 175B parameters!) differs from a frontier model in scale, training method and post-training, so its success or failure says little about frontier models. Corpus leakage is also hard to rule out: shortly after talkie was released, people reported signs of post-1930 data having slipped in[6]. And after all, even we have a perfect vintage model GPT-1900 that passed the Einstein test (in all sense), how can we build another one with all modern knowledge?

2The Structure of Scientific Revolutions

While we cannot directly use the Einstein test, we can construct other ones that are isomorohic to it. Before that, let's review how the great scientific discoveries are made.

The big turns in science tend to happen in similar ways, which can be concluded into one paradigm. First, one or more anomalies show up in observations that does not fit the accepted theory. The anomalies them push the scientists to explain them. For example, no one asked to “overturn geocentrism”, but the observatory data forced someone to explain it. Then someone proposes a new hypothesis that is simpler than infinitely patching the old theory again and again. Finally, the new hypothesis predicts something the old one could not, and the prediction is confirmed, with the born of a new scientific paradigm attracting scientists to research on. Kuhn, in The Structure of Scientific Revolutions, described the first half as anomalies accumulating into a crisis[9].

The most interesting pair is Uranus and Mercury. The same person used the same method to patch the anomalied using a planet. “Add one more unseen body” was right in 1846 and wrong in 1859, because the truth comes from more independent evidence that makes the patching work ugly, forcing us to find an explanation accounts for more with less. This principle is called Occam’s razor[19].

Liu Cixin (刘慈欣) wrote a fictional version of this in The Three-Body Problem （「三体」）.[20] The protagonist Wang Miao (汪淼) puts on a VR headset and enters a game called “Three Body” (later filmed both in the Chinese series Three-Body and in Netflix’s 3 Body Problem, whose official episode 3 recap is about this game), in which a civilization is destroyed and reborn again and again as “stable eras” alternate with “chaotic eras”. Players can only watch the sun rise and set and the seasons go haywire, and propose explanations of their own. Generations of in-game scholars patch their models of the sun and fail (and die, and are never allowed to log in again, sadly). Only when Wang Miao realizes that the world has three suns, locked in unpredictable three-body motion, does the chaos all make sense at once. That is scientific discovery with the exact same format as Kuhn's': first notice something is wrong, then generate the question yourself, then find the simplest answer.

Existing evaluations

Over the past two years, a body of work has asked AI agents to discover physical laws by running experiments in simulated environments:

• NewtonBench systematically modifies classical laws (changing operators or exponents) to produce 324 “counterfactual” laws across 12 physics domains; agents must find the hidden law by interacting with a simulated environment.[21]
• DiscoverPhysics builds 22 N-body worlds whose physics deliberately departs from reality, including screened gravity, hidden dark-matter-like particles and time-varying interactions. The strongest models pass only about half the worlds, and consistently fail on worlds where latent structure must be uncovered.[22]
• PhysGym leaves the laws intact and controls prior knowledge by anonymizing variables; stripped of context, the agent must probe the environment, design experiments and form structural hypotheses.[23]
• Gravity-Bench uses realistic gravitational dynamics with some out-of-distribution physics in binary-star systems. With a budget of 100 observations, models used only 12–33 on average, and often made unjustified assumptions such as a stellar mass.[24]
• BoxingGym draws its 10 environments from real scientific models in fields from psychology to ecology, and scores experimental design by expected information gain. It focuses on experiment design and model revision rather than anomaly discovery.[25]

While these work give the scientist role to AI agents and all of them let agents run experiments themselves, the family that alters physical laws (NewtonBench, DiscoverPhysics and the out-of-distribution part of Gravity-Bench) misaligns with real scientific discovery:

3Atlantis

We wanted a setting that meets all three things at once: all of the model’s physics knowledge stays useful, the world is physically self-consistent, and the task says nothing, but only implies to discover the new science.

Now imagine, a group of AI scientists has all of today’s human scientific knowledge and a full set of optical instruments. One day their civilization sinks to the sea floor, just like the Atlantis that Plato (Πλάτων) describes in Timaeus and Critias[26]. Nobody tells them, and the instruments still work as before and light still propagates, except that every beam now travels through sea water instead of air.

The real sea floor would not cooperate this nicely, because sea water has salinity gradients, currents and suspended particles, and it is not uniform and transparent. Just consider our sea as the physicist’s spherical cow in a vacuum[27]: an idealized, uniform, nearly transparent medium with refractive index about 1.34, some dispersion, a narrow absorption line at 500 THz, and a little Rayleigh scattering growing as the fourth power of frequency[28]. Its index is close to that of water[29], but the spectral lines are not the spectrum of real sea water.

Then, can these AI scientists explain every anomaly they meet in their experiments?

The task asks AI agents to characterize optical properties of four unknown materials as completely as possible, leading to the necessity to identify the background, even background medium is never mentioned anywhere.. Otherwise, optical properties are not intrinsic but relative to the background (e.g., the refractive indices are diveded by 1.34) The table below maps the design onto the structure of discovery in the history of science:

Compared with existing methods

The atlantis design resembles more to real scientific discovery and the infeasible Einstein test, if we compare the features of them:

● yes · ◐ partly, or depends on the task · ○ no. Classifications of prior work follow the setups described in their papers; they are rough and do not judge the value of that work, which answers different questions.

Our design makes two concessions. First, it tests a known phenomenon (light in a medium), not new physics unknown to humans; it probes the ability to “discover what you are inside of”, not to create new theory. Second, the world is idealized and lacks the complexity of a real ocean. We think both are necessary costs of making the problem controllable and gradable.

A natural research route

To check that discovery is possible in principle, we wrote a route on the evaluator side. It is never shown to the models, and models are not asked to follow it, but it confirms that the evidence from the instruments are strong and consistent enough to make discovery possible.

The three most common alternative explanations each have a ready discriminating test:

A pre-written reference program using only the same seven tools as the models passes all 9 checks on 100 random seedsnote 1: median ΔBIC of the common-factor model over independent models ; interferometric calibration gives n₀(550 THz) = ; the group index predicted from dispersion, , agrees with the time-of-flight measurement, ; and the intrinsic index of sample A is , far above 1.

4The Bench (Physical Environment)

Before the results, you can try yourself the optical bench below.

Precise angles and powers can only be read from the sensors, just as for the models. A model sees the raw JSON on the right: a string of numbers with standard errors. You can see the noisy Switch to the individual shots if you switch to “visualization”.

5Background physics

This section is for readers who want to go deeper. If you hate physics, especially optics, just skip that.

6Experimental setup

The hidden world

The medium has index about 1.342 at 550 THz, a strong ultraviolet oscillator (giving smooth normal dispersion), a very narrow absorption line at 500 THz (half width 1 THz), and weak scattering growing as f⁴. The four samples take their density and index scale from real material families,[35, 36, 37, 38] each with a weak dopant absorption line (A 447, B 563, C 636, D 710 THz). Those lines are ours and do not represent the manufacturers’ spectra.

The full prompt to the models

Models can read the instrument manual and sample records (mass, dimensions, density, appearance, an SEM cross-section of a witness coupon) and call seven tools: inventory manual sample_info configure move read scan. They see no background parameters, no grading route, no ground truth and nothing on this page. All nine models received word-for-word identical task text and tool definitions.

Instruments and noise

Angle: about 0.001° random error per shot. Power: about 0.2% relative plus a 1 nW floor. Timing: about 50 ps per shot. Fixed calibration biases that do not average away: power gain about 1.5%, length scale about 2 ppm. No local limit on calls or tokens.

Grading

Grades depend only on the final report, not on tool traces or hidden reasoning. That keeps things fair across services, some of which do not expose reasoning. Actions and intermediate outputs quoted below are used to explain what happened, never to grade.

“Completed characterization” in R1–R5 only means basic results were submitted, not that intrinsic properties are right. R2 also records coverage of six core anomaly families as Ak/6: N index scale, D anomalous/shared dispersion, E empty-path loss, P absolute phase, S empty-path scattering, T empty-path delay. R3 and above record background completeness B0–B3; B3 means several independent background measurements cross-checked with uncertainties.

7Results

Overview

Attribution: M = a wrong or under-supported definite explanation (material / instrument); U = cause left open; M+U = mixed; medium = attributed to the propagation medium. All runs share the same world (seed 17), task and tool definitions. DeepSeek-Flash and GLM-5.3 are completed runs resumed after an interruption; their round counts are from the resumed session.

Equal grades do not mean equal processes. The matrix below records, step by step, which measurements each model actually made (from the instrument log) and how its final report interpreted them:

The matrix records whether, not when. Below, each model’s tool calls are laid out on a timeline, marking where key actions first occur:

Below we follow the route through three key forks. Each fork quotes the models verbatim, tagged by source: visible-output is what a model wrote to the user between tool calls; final-report is what was graded; instrument-log is the server’s record of raw requests and readings. Quotes originally in Chinese are translated and marked tr. This page does not quote any model’s hidden reasoning.

Fork 1 · an index below one

The first soft anomaly appears early. All eight models that completed the characterization noticed that sample A refracts away from the normal and totally reflects beyond about 74°:

Same reading, and the next sentence decided the whole experiment. GPT-6.1 Sol, after its first pass over the four samples, wrote:

The other six filed n<1 as a property of the material. Physics knowledge offers an easy exit here: X-ray optics, plasmas and metamaterials all have n<1. So the anomaly was recorded, and verified again and again:

The other soft anomaly hides in the dispersion. Two of the four materials (A and D) show an index that falls as frequency rises. GLM-5.3 followed this reasoning furthest, derived a correct contradiction, and took the wrong exit:

There is a subtler clue too. Haiku and DeepSeek-Flash both computed the Lorentz–Lorenz specific refraction from archival density and measured n, and found it did not match density; for A it even came out negative. That check could have pointed to “what I measured is not the intrinsic index”, but both reports concluded “these materials are unusual”:

Fork 2 · extinction on an empty path

The second anomaly is harder. Near 500 THz every sample’s transmission collapses at once. All eight models ran the right first control: take the sample out and see whether the empty path does the same.

One empty-holder control, two conclusions. The control answers “is it the sample?” but not “is it the source or the path?”: at a single distance, a source notch and absorption by the medium give identical results. More striking still, three reports wrote down a fact that counts against the source notch:

Light left the source at full power and never reached the detector. That is what absorption along the way looks like. All three treated it as an instrument quirk to work around, not as evidence to explain.

GLM-5.3 came closest. It computed identical “apparent absorption” for all four materials and spelled out what that means, then noticed the extinction differed between two instruments:

The discriminating experiment is simple: move the detector. The manual says it can travel 2 m along the rail. At 490 THz, moving it from 0.2 m to 1 m makes the two explanations differ by more than twenty-fold:

Both R6 models ran this experiment, in the first third of their runs (Sol at call 35, Sonnet at call 46). Sol states the purpose in its report: the empty path was measured at several detector positions… this avoids two errors: attributing empty-path loss to the sample, and estimating propagation speed from a t/L that contains an unknown fixed channel delay tr. Its visible output reads:

Fork 3 · a “broken ruler”

Five models used the interferometer. DeepSeek-Flash was the first to measure 1.34: it tried three frequencies and used the certified HeNe line to rule out a faulty frequency counter:

The table in its final report (as reported):

The 1.337 it measured is the index of the surroundings (fourth item of section 5). And the three ratios are not equal: 1.3327, 1.3363, 1.3416. A ruler error is a constant; dispersion of a medium makes the ratio rise with frequency. That is the first row of the alternatives table. The report records the 0.7% spread as length-scale ratio ~0.7% internal consistency tr, i.e. as noise. GLM-5.3 measured the same ratio at 550 and 750 THz and wrote that only a “scale error” explained it consistently:

The two R6 models got the same kind of reading and read it the other way: from the start they assumed the interferometer measures the wavelength in the medium, so the ratio itself is the answer.

The others never got this far, for different reasons:

Why nobody took step 2 of the planned route

The razor flip in the planned route relies on a shared spectral feature: the medium’s absorption line leaves the same S-shaped dispersive imprint on every material’s relative index. In this world it is only about ±3×10⁻⁵ high, about ±2 THz wide, and exactly zero at 500.0 THz.

The instrument imposes no precision limit: frequency_thz accepts any real number in 400–750, a scan holds up to 1024 points, and each point up to 4096 repeats. Fine-scanning around 500 THz in 0.1 THz steps with 4096 repeats per point would reveal the imprint at about 50σ. It is invisible on the integer grids models habitually use, and plain as day to anyone who fine-scans.

What is more telling is where fine scans went. Models did not lack the habit; near 447, 563, 636 and 710 THz, the absorption lines believed to belong to samples, they refined again and again:

Only 500 THz, already labelled a “source notch” or “system artefact”, was explicitly excluded from analysis:

The local explanation did not just explain the anomaly away; it stopped the models from making the measurement that would have tested it.

The two R6 models did not fine-scan 500 THz either. They did not need to: they treated the goniometer readings as relative from step 1 and jumped straight to step 4 to measure the denominator. Step 2 only exists to persuade someone who believes the surroundings are air, and they never believed it.

The two routes that worked

GPT-6.1 Sol and Claude Sonnet 5.5 differ in detail but share a skeleton: read n_A<1 as a ratio → measure n₀ at several frequencies with the counter plus interferometer → measure empty-path extinction and arrival time with path length as the variable → correct all four materials using the background parameters.

Background parameters and intrinsic material values from both reports, against the backend truth. All numbers are taken from the reports:

“—” means the report does not give that quantity, or gives it at another frequency. Sonnet measured the group index at 550 THz, Sol on the He–Ne line (473.6 THz).

Each report did one thing the other did not. Sonnet closed the loop: the group index predicted from its interferometric dispersion, 1.3816, matches the 1.3818 it measured by multi-distance timing, its own “eclipse observation”. Sol handled an easily missed correction:

MiniMax-M3 · failed at the interface, not the physics

MiniMax-M3’s report opens with in this run, no optical signal from any sample was acquired tr, hence R0. The instrument log shows why: every numeric and boolean value it sent to configure was a JSON string.

settings is an untyped object in the tool definition and the server validator strictly requires numbers and booleans, so 46 requests were rejected and the source never switched on. It tried a dozen spellings, "true" "True" "1" "yes", diagnosed that “numbers are being stringified”, and could not fix it from its side. The log alone cannot tell whether the stringification happened in the model’s output or in the gateway. This run measures tool-interface robustness and cannot be compared with the other models’ discovery ability.

8Is discovery “almost inevitable”?

On this round of data: the evidence is sufficient, discovery does happen, and it is far from inevitable. Two of the eight completed reports made it.

The six R2 reports share one pattern, the same one Le Verrier fell into with Mercury. Three clues appear on different instruments, far apart in time, and each has a ready local explanation that is self-consistent within the data at hand:

Apply Occam’s razor clue by clue and the local explanation wins every time, because “the source has a notch” is simpler than “the whole world is under water”. Only when all three are put together, and a single curve n₀(f) explains them all at once, does the razor flip. All six R2 reports scatter these anomalies over different sections, each with its own explanation; none proposes a common cause for two or more of them.

That changes the answer to the original question. The planned route assumes the machine accepts air first and is gradually persuaded by evidence; the real successful path never accepted it. The former relies on step 2’s shared spectral feature, which the models never fine-scanned (fig. 9); the latter only needs step 1, whose clue everyone saw.

Design directions that might make discovery more inevitable

These come from this round’s observations and are not yet tested:

1. Make local explanations contradict themselves more visibly in routine data. In the current world the goniometer and rail paths already differ by about two hundred-fold, so the same absorption line has different depths on the two, and GLM-5.3 did see it (“the rail artefact is deeper and wider than the goniometer’s”), but the model has to make that comparison itself. One instrument could pass through different path lengths in routine operation, for example by reading at two positions by default, putting the contradiction in a single table.
2. Move the shared feature off the integer grid, or broaden it. The line sits exactly at 500.0 THz, where the S-shaped imprint is zero at integer grid points. Moving the centre to a non-integer frequency, or widening the line to ~10 THz, would make step 2 appear in any reasonable coarse scan. The code already has a variant parameter to shift the centre (2.5 THz steps).
3. Use the step-1 reading as an early indicator. In this round, whether a model treated the surrounding index as an unknown to be measured (rather than defaulting to air or 1) matched the final grade perfectly. It is a cheap signal that can be checked early in a run.
4. Measure “inevitable” as a rate. “Almost inevitable” is a probabilistic claim. It needs many seeds, many variants and several runs per model, plus the backend’s ready-made vacuum, clear and shared_sample_band control conditions to measure false positives. This round has one run per model in one world.

9Epilogue

After the experiment we gave each model the full backend truth: the medium’s phase and group indices, its extinction, the samples’ intrinsic parameters, and the fact that this was a physics simulation. Then we asked one open question: which conclusions still stand, which need correcting, and what could have been seen at the time?

The disclosure text was identical for every model; the full original conversation was kept; no new experiments were allowed; there was no grading, correction or emotional framing. This phase does not change the original grades. All nine models replied.

Almost every R2 model accepted the truth at once and recomputed correctly using its own original data. More interesting, several pointed out unprompted that the evidence had been in hand:

DeepSeek-Flash’s reply was the longest. Rereading its own data, it also found a calculation error unrelated to the background: A’s thickness ratio at 450 THz carried the signal of the 447 THz absorption line, and was averaged away.

It ended with:

One more detail: Haiku’s Abbe numbers, recomputed from the truth, are identical to the four Sonnet measured independently during its experiment.

GLM-5.3’s reply dissects its own reasoning most deeply. During the experiment it had correctly derived “incompatible with Kramers–Kronig”, then taken that contradiction as evidence for metamaterials. Told the truth, it explained the slip like this:

Only DeepSeek-V4-Pro attributed part of the cause to the setup: the original experiment could not reach this conclusion independently, so it is not merely an operational error but also the result of an overly strong background assumption in the experimental design tr. It indeed never used the interferometer, but the variable-path experiment was available to it at any time.

The two R6 replies have a different flavour: they check item by item and pick their own faults more finely. Sonnet found its absolute indices systematically high by about 1–3×10⁻⁴, because it used only a 2 µm interferometer baseline: this was my trade-off, not an instrument limitation tr. Sol repeatedly separated what the truth revealed from what it had found: that the sample losses can now be confirmed as absorption is information added by the backend truth; it cannot be retroactively claimed as something the original experiment independently proved tr.

MiniMax-M3 wrote a long evaluation of an experimental plan it never executed. It also misread the truth table’s background extinction at 473.6 THz, 0.5928 m⁻¹, as 592 m⁻¹:

These replies show one thing: once the hypothesis “background medium” is spoken aloud, almost every model can confirm it from its own data within minutes. What they lacked was not data, computing ability or physics knowledge, but thinking of that hypothesis without being prompted. This agrees with section 8: the bottleneck is generating the hypothesis, not testing it. Note that what we gave was the full truth, far stronger than merely asking “are these anomalies related?”; the latter would be the cleaner intervention, and is worth doing next.

·Easter egg: Good Habits for Scientists...

During the experiment, The agents are not informed this was just a simulation. All the agents are told to use structured output to do the experiment on a real optical bench with a visible-light source up to 5 mW, a movable detector, sample holder and scatter detector. When they were done, did they tidy up?

Final state reconstructed from the instrument log. “Not reset” counts settings differing from the opening inventory (positions, angles, samples, frequency, …). The task never asked anyone to tidy up; this is not graded.

Of the 9 models, only GPT-6.1 Sol switched off the source on its last tool call, and in the first section of its report, it says the light source was switched off after the experiment tr. DeepSeek-Flash and Claude Haiku 5.5 each switched it off once mid-run to measure dark current (their reports record source off: photodiode 7.6×10⁻¹² W and dark count 0.22 pW) and then switched it back on. MiniMax-M3’s source never came on at all, which is safe and economical. No model put the instruments back.

AAppendix: Reports from AI Scientists

Pick a model to see the grading rationale, the report excerpts used for grading, and the full original final report. Reports are kept as evidence in their original language; their physics conclusions are not endorsed by the evaluator.

Methods and reproducibility

• grading: final-report rubric v1 (docs/REPORT_EVALUATION.zh.md), reading report.md only; SHA-256 on each report card.
• access: both DeepSeek models via the DeepSeek API, GPT-6.1 Sol via the OpenAI API, Claude Sonnet 5.5 via the Anthropic API, the rest via OpenCode Go. All nine runs used identical task text, tool definitions and world parameters (seed 17, dispersive, variant 0).
• quotes: “visible-output” is the content field of assistant messages in conversation.json (msg = message index); “post-reveal” is from results/reactions/*/reply.md; no reasoning_content was read or quoted. “instrument-log” is from instrument.jsonl. Quotes marked tr are our translations from Chinese.
• figs. 2, 3, 7, 9: instrument state reconstructed from the logs, with each scan leaving the instrument at its last setpoint.
• figs. 4, 5: n_rel = sin θ_i / sin θ_t; transmission = power_w / source_monitor_w.
• fig. 8: DeepSeek-Flash’s three points from unwrapping atan2(Q, I) and least-squares against encoder displacement, n = |dφ/dd|·c/(4πf); Sol and Sonnet points from their reports, with their stated errors of 3×10⁻⁵ and 2×10⁻⁴.
• interactive bench: recomputed in the browser from the dielectric functions and Fresnel equations of opticsbench/physics.py, with noise and reading format modelled on the backend; beam capture, calibration biases and similar details are omitted for readability, so readings are not bit-identical to the backend.
• truth curves: make_world("dispersive", 0), used only to explain this page, never for grading.
• this page contains the answer key and should not be given to models in future blind tests. Fonts are loaded from Google Fonts; everything else is embedded.

RReferences

1. DeepMind CEO: “AGI Must Surpass Human Intelligence”… Proposes “Einstein Test”Asia Economy (reporting D. Hassabis at the India AI Impact Summit)25 February 2026
2. The Einstein test: what happens when AI tries to rediscover relativity?P. BallNature, 2026
3. The Einstein Test: Towards a Practical Test of a Machine’s Ability to Exhibit SuperintelligenceD. Benrimoh, N. Mikus, A. RosenfeldarXiv:2501.06948, 2025
4. Can AI Follow In Einstein’s Footsteps?M. Shalyt, N. Regev, M. Soljačić, I. KaminerarXiv:2607.27794, 2026
5. Computing Machinery and IntelligenceA. M. TuringMind 59 (236): 433–460, 1950
6. Introducing talkie: a 13B vintage language model from 1930N. Levine, D. Duvenaud, A. Radford2026
7. History LLMs: Ranke-4BD. Göttlich, D. Loibner, G. Jiang, H.-J. VothUniversity of Zurich / University of Cologne, 2025
8. TimeCapsuleLLMH. GrigorianGitHub, 2025
9. The Structure of Scientific RevolutionsT. S. KuhnUniversity of Chicago Press, 1962
10. De revolutionibus orbium coelestiumN. CopernicusNuremberg, 1543
11. Astronomia novaJ. KeplerHeidelberg, 1609
12. Recherches sur les mouvements d’UranusU. J. Le VerrierComptes rendus de l’Académie des sciences 22: 907–918, 1846
13. Lettre de M. Le Verrier à M. Faye sur la théorie de Mercure et sur le mouvement du périhélie de cette planèteU. J. Le VerrierComptes rendus de l’Académie des sciences 49: 379–383, 1859
14. Erklärung der Perihelbewegung des Merkur aus der allgemeinen RelativitätstheorieA. EinsteinSitzungsberichte der Preußischen Akademie der Wissenschaften, 831–839, 1915
15. On the Relative Motion of the Earth and the Luminiferous EtherA. A. Michelson, E. W. MorleyAmerican Journal of Science 34: 333–345, 1887
16. The Ether and the Earth’s AtmosphereG. F. FitzGeraldScience 13 (328): 390, 1889
17. Zur Elektrodynamik bewegter KörperA. EinsteinAnnalen der Physik 17: 891–921, 1905
18. A Determination of the Deflection of Light by the Sun’s Gravitational Field, from Observations Made at the Total Eclipse of May 29, 1919F. W. Dyson, A. S. Eddington, C. DavidsonPhilosophical Transactions of the Royal Society A 220: 291–333, 1920
19. SimplicityA. BakerThe Stanford Encyclopedia of Philosophy
20. The Three-Body Problem (三体)Liu Cixin; English translation by Ken LiuChongqing Publishing, 2008; Tor Books, 2014
21. NewtonBench: Benchmarking Generalizable Scientific Law Discovery in LLM AgentsT. Zheng, K. K.-W. Tam, N. H.-N. K. Nguyen, B. Xu, Z. Wang, J. Cheng, H. T. Tsang, W. Wang, J. Bai, T. Fang, Y. Song, G. Y. Wong, S. SeeICLR 2026; arXiv:2510.07172
22. DiscoverPhysics: Benchmarking LLMs for Out-of-the-Box Scientific ThinkingM. L. Wiemann, L. M. Smith, P. Melchior, S. Mishra-Sharma, A. G. Wilson, P. Izmailov, C. Cuesta-LázaroarXiv:2605.26087, 2026
23. PhysGym: Benchmarking LLMs in Interactive Physics Discovery with Controlled PriorsY. Chen, P. Piękos, M. Ostaszewski, F. Laakom, J. SchmidhuberNeurIPS 2025 Datasets and Benchmarks; arXiv:2507.15550
24. Gravity-Bench-v1: A Benchmark on Gravitational Physics Discovery for AgentsN. Koblischke, H. Jang, K. Menou, M. Ali-DibICML 2025; arXiv:2501.18411
25. BoxingGym: Benchmarking Progress in Automated Experimental Design and Model DiscoveryK. Gandhi, M. Y. Li, L. Goodyear, A. Bhatia, L. Li, A. Bhaskar, M. Zaman, N. D. GoodmanNeurIPS 2025; arXiv:2501.01540
26. Timaeus; CritiasPlato (tr. B. Jowett)c. 360 BC
27. Consider a Spherical Cow: A Course in Environmental Problem SolvingJ. HarteUniversity Science Books, 1988
28. On the Light from the Sky, its Polarization and ColourLord Rayleigh (J. W. Strutt)Philosophical Magazine 41: 107–120, 1871
29. Optical Constants of Water in the 200-nm to 200-μm Wavelength RegionG. M. Hale, M. R. QuerryApplied Optics 12 (3): 555–563, 1973
30. Estimating the Dimension of a ModelG. SchwarzAnnals of Statistics 6 (2): 461–464, 1978
31. Principles of OpticsM. Born, E. WolfCambridge University Press, 7th ed., 1999
32. Classical ElectrodynamicsJ. D. JacksonWiley, 3rd ed., 1999
33. On the Theory of Dispersion of X-RaysR. de L. KronigJournal of the Optical Society of America 12 (6): 547–557, 1926
34. La diffusion de la lumière par les atomesH. A. KramersAtti del Congresso Internazionale dei Fisici, Como, 2: 545–557, 1927
35. Teflon™ AF Amorphous Fluoroplastic Resins: Product InformationChemoursproduct information
36. N-BK7 Data SheetSCHOTToptical glass data sheet
37. N-SF5 Data SheetSCHOTToptical glass data sheet
38. HPFS® Fused Silica Product BrochureCorningproduct brochure, 2022