arXiv subjects

Sandeep Varma

Publications and source records attributed to Sandeep Varma.

6 recordsLinked to original sources

arXiv2026

Aryabhata 2: Scaling Reinforcement Learning for Advanced STEM Reasoning

Competitive STEM examinations such as JEE and NEET require multi-step symbolic reasoning, precise numerical computation, and deep conceptual understanding across physics, chemistry, and mathematics. Recent large language models perform strongly on common reasoning benchmarks, yet they remain difficult to deploy at scale, where millions of student doubts demand domain-specific, consistently structured problem solving. We introduce Aryabhata 2, a reasoning-focused language model for competitive STEM examinations, trained via reinforcement-learning post-training. Using PhysicsWallah's internal question banks, we construct a high-quality training curriculum and post-train GPT-OSS-20B through reinforcement learning with verifiable rewards. Training combines prolonged reinforcement learning with broadened exploration via progressively larger rollout group sizes. We evaluate Aryabhata 2 on competitive examination benchmarks, including JEE Main, JEE Advanced, and NEET, as well as out-of-distribution reasoning datasets such as AIME, HMMT, MMLU-Pro, MMLU-Redux 2.0, and GPQA. Results show that Aryabhata 2 outperforms its base model GPT-OSS-20B on competitive STEM reasoning while requiring substantially fewer output tokens (up to 64\% fewer).

cs.CL↗

arXiv2025

Aryabhata: An exam-focused language model for JEE Math

We present Aryabhata 1.0, a compact 7B parameter math reasoning model optimized for the Indian academic exam, the Joint Entrance Examination (JEE). Despite rapid progress in large language models (LLMs), current models often remain unsuitable for educational use. Aryabhata 1.0 is built by merging strong open-weight reasoning models, followed by supervised fine-tuning (SFT) with curriculum learning on verified chain-of-thought (CoT) traces curated through best-of-$n$ rejection sampling. To further boost performance, we apply reinforcement learning with verifiable rewards (RLVR) using A2C objective with group-relative advantage estimation along with novel exploration strategies such as Adaptive Group Resizing and Temperature Scaling. Evaluated on both in-distribution (JEE Main 2025) and out-of-distribution (MATH, GSM8K) benchmarks, Aryabhata outperforms existing models in accuracy and efficiency, while offering pedagogically useful step-by-step reasoning. We release Aryabhata as a foundation model to advance exam-centric, open-source small language models. This marks our first open release for community feedback (https://huggingface.co/PhysicsWallahAI/Aryabhata-1.0); PW is actively training future models to further improve learning outcomes for students.

cs.AI↗

arXiv2024

On congruent isomorphisms for tori

Let $F$ and $F'$ be two $l$-close nonarchimedean local fields, where $l$ is a positive integer, and let $\mathrm{T}$ and $\mathrm{T}'$ be two tori over $F$ and $F'$, respectively, such that their cocharacter lattices can be identified as modules over the ''at most $l$-ramified'' absolute Galois group $Γ_F/I_F^l \congΓ_{F'}/I_{F'}^l$. In the spirit of the work of Kazhdan and Ganapathy, for every positive integer $m$ relative to which $l$ is large, we construct a congruent isomorphism $\mathrm{T}(F)/\mathrm{T}(F)_m\cong\mathrm{T}'(F')/\mathrm{T}'(F')_m$, where $\mathrm{T}(F)_m$ and $\mathrm{T}(F')_m$ are the minimal congruent filtration subgroups of $\mathrm{T}(F)$ and $\mathrm{T}(F')$, respectively, defined by J.-K.~Yu. We prove that this isomorphism is functorial and compatible with both the isomorphism constructed by Chai and Yu and the Kottwitz homomorphism for tori. We show that, when $l$ is even larger relative to $m$, it moreover respects the local Langlands correspondence for tori.

math.NT↗

arXiv2020

Learning semantic Image attributes using Image recognition and knowledge graph embeddings

Extracting structured knowledge from texts has traditionally been used for knowledge base generation. However, other sources of information, such as images can be leveraged into this process to build more complete and richer knowledge bases. Structured semantic representation of the content of an image and knowledge graph embeddings can provide a unique representation of semantic relationships between image entities. Linking known entities in knowledge graphs and learning open-world images using language models has attracted lots of interest over the years. In this paper, we propose a shared learning approach to learn semantic attributes of images by combining a knowledge graph embedding model with the recognized attributes of images. The proposed model premises to help us understand the semantic relationship between the entities of an image and implicitly provide a link for the extracted entities through a knowledge graph embedding model. Under the limitation of using a custom user-defined knowledge base with limited data, the proposed model presents significant accuracy and provides a new alternative to the earlier approaches. The proposed approach is a step towards bridging the gap between frameworks which learn from large amounts of data and frameworks which use a limited set of predicates to infer new knowledge.

cs.CV↗

arXiv2020

The Bernstein projector determined by a weak associate class of good cosets

Let $G$ be a reductive group over a $p$-adic field $F$ of characteristic zero, with $p \gg 0$. In [Kim04], J.-L. Kim studied an equivalence relation called weak associativity on the set of unrefined minimal $K$-types for $G$ in the sense of A. Moy and G. Prasad. Following [Kim04], we attach to the set $\overline{\mathfrak s}$ of good $K$-types in a weak associate class of positive-depth unrefined minimal $K$-types a $G(F)$-invariant open and closed subset $\mathfrak g(F)_{\overline{\mathfrak s}}$ of the Lie algebra $\mathfrak g(F)$ of $G(F)$, and a subset $\tilde G_{\overline{\mathfrak s}}$ of the admissible dual $\tilde G$ of $G(F)$ consisting of those representations containing an unrefined minimal $K$-type that belongs to $\overline{\mathfrak s}$. Then $\tilde G_{\overline{\mathfrak s}}$ is the union of finitely many Bernstein components for $G$, so that we can consider the Bernstein projector $E_{\overline{\mathfrak s}}$ that it determines. We show that $E_{\overline{\mathfrak s}}$ vanishes outside the Moy--Prasad $G(F)$-domain $G(F)_r \subset G(F)$, and reformulate a result of Kim as saying that the restriction of $E_{\overline{\mathfrak s}}$ to $G(F)_r$, pushed forward via the logarithm to the Moy--Prasad $G(F)$-domain $\mathfrak g(F)_r \subset \mathfrak g(F)$, agrees on $\mathfrak g(F)_r$ with the inverse Fourier transform of the characteristic function of $\mathfrak g(F)_{\overline{\mathfrak s}}$. This is a variant of one of the descriptions given by R. Bezrukavnikov, D. Kazhdan and Y. Varshavsky in arXiv:1504.01353 for the depth-$r$ Bernstein projector.

math.RT↗

arXiv2018

On Kostant Sections and Topological Nilpotence

Let G denote a connected, quasi-split reductive group over a field F that is complete with respect to a discrete valuation and that has a perfect residue field. Under mild hypotheses, we produce a subset of the Lie algebra g(F) that picks out a G(F)-conjugacy class in every stable, regular, topologically nilpotent conjugacy class in g(F). This generalizes an earlier result obtained by DeBacker and one of the authors under stronger hypotheses. We then show that if F is p-adic, then the characteristic function of this set behaves well with respect to endoscopic transfer.

math.RT↗