TheoPhysEval: Assessing AI agents at autonomous theoretical physics problem solving

A. Ramesh Chandra; Aaghaz Mahajan; Ainesh Sanyal; Mianqi Wang; Shubhadeep Chakraborty; Sudip Chakraborty; Shailesh Chandrasekharan; Rama Govindarajan; Akshit Goyal; Jonathan J. Heckman; Yasir Iqbal; Divya Jaganathan; Saumav Kapoor; Andreas Karch; Pushkar Khandare; Sandeep Krishna; Anupam Kundu; Alok Laddha; Tongyan Lin; Swarnendu Maity; Nilmani Mathur; Shani Nadir Meynet; Alessandro Mininno; Abhijeet Mishra; Vidur Niranjan; Abhishodh Prakash; Pratik Rath; Durbar Roy; Debargha Sarkar; Srikanth Sastry; Gary Shiu; Aninda Sinha; Thomas Van Riet; Praneeth Netrapalli; Suvrat Raju

arXiv · 2610.16733

TheoPhysEval: Assessing AI agents at autonomous theoretical physics problem solving

View original source ↗Read source document ↗Export metadata ↓

Abstract

We present a suite of 50 unpublished research-level theoretical physics problems with verifiable answers. Without AI assistance, a typical problem in our collection could be solved by a world expert in about a day or less but would take considerably longer for experts in adjacent specialties. We test the ability of off-the-shelf AI agents to tackle these problems autonomously, without human hints. When given access to tools available in a realistic setting, including the ability to search the literature and write scripts for numerical and symbolic work, the best-performing agent, GPT-6-Astra harnessed in Codex, obtains the final answers with a near-perfect accuracy of 98% at an average cost under USD 3 per problem. On average, agents show only small changes in accuracy when denied web search and access to background research papers, suggesting significant memorization of the literature. These results suggest that state-of-the-art agents can accurately perform concrete calculations required across theoretical physics research. We comment on the broader implications of these findings and discuss possible explanations for the current lack of agent-led breakthroughs in theoretical physics despite their outstanding problem-solving abilities.

Explore related subjects

hep-th cond-mat.stat-mech cond-mat.str-el gr-qc

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

A. Ramesh Chandra, Aaghaz Mahajan, Ainesh Sanyal, Mianqi Wang, Shubhadeep Chakraborty, Sudip Chakraborty, Shailesh Chandrasekharan, Rama Govindarajan, Akshit Goyal, Jonathan J. Heckman, Yasir Iqbal, Divya Jaganathan, Saumav Kapoor, Andreas Karch, Pushkar Khandare, Sandeep Krishna, Anupam Kundu, Alok Laddha, Tongyan Lin, Swarnendu Maity, Nilmani Mathur, Shani Nadir Meynet, Alessandro Mininno, Abhijeet Mishra, Vidur Niranjan, Abhishodh Prakash, Pratik Rath, Durbar Roy, Debargha Sarkar, Srikanth Sastry, Gary Shiu, Aninda Sinha, Thomas Van Riet, Praneeth Netrapalli, Suvrat Raju. 2026-10-09. TheoPhysEval: Assessing AI agents at autonomous theoretical physics problem solving. https://arxiv.org/abs/2610.16733

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Groupoid Representation of Lorentz Transformations Including Superluminal Observers

We construct a Minkowski-pointed Poincare groupoid that extends the geometric description of Lorentz transformations to frames superluminal relative to a distinguished rest frame. Real four-dimensional superluminal boost matrices factorize into ordinary Lorentz boosts and finite infinite-speed limits. For each spatial direction, the two oriented limits define arrows between the Minkowski metric object and a directional metric object; their composition yields an involutive isotropy loop generating a direction-dependent extension of the Poincare group. These arrows induce *-isomorphisms of Poincare observable algebras, transporting generators and their tensor products. Since all metric objects share a common affine spacetime and dual momentum space, transported spectral data can be classified relative to the distinguished Minkowski metric. This Minkowski-pointed classification recovers ordinary massive and massless Wigner sectors and assigns spacelike orbits, with SO(2,1) little groups, to massive momentum representatives transported to directional objects. It thus provides a geometric interpretation of tachyonic Wigner wavefunction classes without postulating intrinsically tachyonic states. As a kinematic application, we examine polarized cobalt-60 beta decay. Under an explicit source-arrow correspondence, reversal of nuclear polarization exchanges oriented arrows and reverses the preferred momentum direction, consistently with the observed reversal of the Wu counting asymmetry. The measured angular law is compatible with this correspondence; a derivation of decay amplitudes and asymmetry coefficients remains open.

hep-th↗

arXiv2026

Timelike Holographic Complexity

Motivated by the pseudo-entropy program, we study timelike subregion complexity within the holographic Complexity=Volume framework, extending previous spatial constructions to Lorentzian boundary intervals. For hyperbolic timelike regions in pure AdS, we compute the enclosed bulk volume and show that, despite the Lorentzian embedding, the resulting complexity is purely real under the stated volume prescription. We generalize the analysis to AdS black brane geometries, where the extremal-surface configurations considered may either remain outside the horizon or penetrate it, with the horizon-to-turning-point segment being timelike in the exterior case and spacelike in the interior case. In both configurations, the complexity exhibits the same leading UV-divergence structure as the spacelike case and receives no imaginary contribution when the horizon terms are combined with a common regulator. This reality contrasts with the potentially complex-valued pseudo-entropy and reflects the chosen Lorentzian volume prescription, while a precise boundary interpretation of the proposed complexity remains conjectural.

hep-th↗

arXiv2026

A Master Superspace Action for 3D S-Duality

We formulate a `master' partition function in three-dimensional $\mathcal{N}=2$ superspace that realises, upon integrating out complementary superfields, both the electric Maxwell--Chern--Simons (MCS) theory and its magnetic $S$-dual: a non-gauge Deser--Jackiw self-dual massive vector times a decoupled level-$k$ Chern--Simons term. The two descriptions share the topological mass $M=\frac{g^{2}k}{2π}$ and obey an exact partition-function identity $\mathcal Z_{\rm mag}(g_m^2,k)=\mathcal Z_{\rm ele}(g_e^2,k)$ with $g_e g_m=2π$, mapping a weakly coupled MCS theory to a strongly coupled Deser--Jackiw CS theory. Special limits reproduce pure Chern--Simons/Gaiotto--Witten ($g^{2}=0$) and Maxwell/compact-scalar duality ($k=0$). We extend the construction to a non-Abelian U$(N)$ gauge group obtaining $\mathcal N=2$ Yang--Mills--Chern--Simons on the electric side and a massive non-gauge vector coupled to level-$k$ Chern--Simons on the magnetic side; the interaction terms between the massive vector and the Chern-Simons term vanish in the Abelian case. Decomposing the $\mathcal N=2$ vector into an $\mathcal N=1$ vector and a real-linear multiplet factorises the master action and yields the $\mathcal N=1$ counterparts. This uplifts the bosonic duality formulated recently to $\mathcal N=2$ and clarifies its non-Abelian and $\mathcal N=1$ reductions.

hep-th↗