arXiv subjects

Piyush Jha

Publications and source records attributed to Piyush Jha.

13 recordsLinked to original sources

arXiv2026

CDRL: Certification-Driven Reinforcement Learning for Neutrino Flavor Model Discovery

Many scientific discovery problems require searching combinatorial hypothesis spaces under complex domain constraints. Reinforcement learning (RL) offers a promising approach, but existing methods rely on scalar rewards that provide limited information about why candidate solutions fail, leading agents to repeatedly explore invalid regions. We introduce Certification-Driven Reinforcement Learning (CDRL), a framework that leverages structured feedback from symbolic reasoning tools. When a candidate violates domain constraints, these tools produce certificates identifying the actions responsible for failure. CDRL converts these certificates into reusable constraints that eliminate classes of invalid solutions and guide exploration toward valid regions. We evaluate CDRL on neutrino flavor model discovery in theoretical particle physics, where the hypothesis space exceeds $10^{26}$ possible models, and compare it with the state-of-the-art RL approach previously used for this task. Across three theory spaces, CDRL achieves up to 1.95$\times$ higher valid model rates and up to 6.33$\times$ higher neutrino model rates while evaluating up to 4$\times$ fewer candidates. We further extract 40 interpretable rules from search trajectories using a post-hoc decision-tree framework and show that reusing them as soft constraints yields gains of up to 2$\times$ in valid model rates and 3$\times$ in neutrino model discovery across all three theory spaces. These results suggest that CDRL uncovers reusable structure in combinatorial search spaces and provides a general framework for scientific model discovery.

cs.AI↗

arXiv2026

MathConstraint: Automated Generation of Verified Combinatorial Reasoning Instances for LLMs

We introduce MathConstraint, a hard, adaptive benchmark for evaluating the combinatorial reasoning capabilities of LLMs. We combine constraint satisfaction problems with rigorous solver-based verification and design an adaptive generator to create instances that remain challenging as the LLMs improve in their reasoning capabilities. Unlike existing benchmarks that quickly saturate on fixed datasets or use LLM-as-a-judge for checking solutions,MathConstraint uses parameterized problem types that enable scalable generation of arbitrarily difficult and automatically verifiable instances. We release MathConstraint-Easy ($266$ instances), on which frontier models achieve between $72.6\%$ (gemini-3.1-flash-lite) and $87.6\%$ (gpt-5.5) accuracy, and MathConstraint ($329$ instances) on which the same models drop to between $18.5\%$ (claude-4.6-sonnet) and $66.9\%$ (gpt-5.5) accuracy, demonstrating the resilience of our benchmark generator against rapid progress in LLM reasoning capabilities. We evaluate 12 frontier and open-weight models with and without access to a sandboxed Python environment that includes generic SAT/SMT solvers. Tool access roughly doubles frontier accuracy on MathConstraint (mean $+28$pp; up to $+52$pp for claude-4.6-sonnet). Further, halving the tool-call budget from $8$ to $4$ rounds erases up to $37$ points -- a sensitivity that most single-budget benchmarks miss. We release the generator, dataset, and evaluation harness as a robust environment for studying combinatorial reasoning and tool-use behavior under adversarially-tunable difficulty.

cs.LG↗

arXiv2026

LLMStinger: Jailbreaking LLMs using RL fine-tuned LLMs

We introduce LLMStinger, a novel approach that leverages Large Language Models (LLMs) to automatically generate adversarial suffixes for jailbreak attacks. Unlike traditional methods, which require complex prompt engineering or white-box access, LLMStinger uses a reinforcement learning (RL) loop to fine-tune an attacker LLM, generating new suffixes based on existing attacks for harmful questions from the HarmBench benchmark. Our method significantly outperforms existing red-teaming approaches (we compared against 15 of the latest methods), achieving a +57.2% improvement in Attack Success Rate (ASR) on LLaMA2-7B-chat and a +50.3% ASR increase on Claude 2, both models known for their extensive safety measures. Additionally, we achieved a 94.97% ASR on GPT-3.5 and 99.4% on Gemma-2B-it, demonstrating the robustness and adaptability of LLMStinger across open and closed-source models.

cs.LG↗

arXiv2026

AlphaMapleSAT: An MCTS-based Cube-and-Conquer SAT Solver for Hard Combinatorial Problems

This paper introduces AlphaMapleSAT, a Cube-and-Conquer (CnC) parallel SAT solver that integrates Monte Carlo Tree Search (MCTS) with deductive feedback to efficiently solve challenging combinatorial SAT problems. Traditional lookahead cubing methods, used by solvers such as March, limit their search depth to reduce overhead often resulting in suboptimal partitions. By contrast, AlphaMapleSAT performs a deeper MCTS search guided by deductive rewards from SAT solvers. This approach enables informed exploration of the cubing space while keeping cubing costs low. We demonstrate the efficacy of our technique via extensive evaluations against the widely used and established March cubing solver on three well-known challenging combinatorial benchmarks, including the minimum Kochen-Specker (KS) problem from quantum mechanics, the Murty-Simon Conjecture, and the Ramsey problems from extremal graph theory. We compare AlphaMapleSAT against March using different types of conquering solvers such as SAT Modulo Symmetries (SMS) and SAT+CAS, both built on top of the CaDiCaL SAT solver. We show that in all cases, there is a speedup in elapsed real time (wall clock time) ranging from 1.61x to 7.57x on a 128 core machine for the above-mentioned problems. We also perform cube-level and parallel scaling analysis over 32, 64, and 128 cores, which shows that AlphaMapleSAT outperforms March on all these settings. Our results show that deductively-guided MCTS search technique for cubing in CnC solvers can significantly outperform March on hard combinatorial problems.

cs.AI↗

arXiv2025

RLSF: Fine-tuning LLMs via Symbolic Feedback

Large Language Models (LLMs) have transformed AI but often struggle with tasks that require domain-specific reasoning and logical alignment. Traditional fine-tuning methods do not leverage the vast amount of symbolic domain-knowledge available to us via symbolic reasoning tools (e.g., provers), and are further limited by sparse rewards and unreliable reward models. We introduce Reinforcement Learning via Symbolic Feedback (RLSF), a novel fine-tuning paradigm where symbolic reasoning tools (e.g., solvers, provers, and algebra systems) provide fine-grained feedback to LLMs. RLSF uses poly-sized certificates (e.g., proofs) generated by symbolic tools to identify and correct errors in model outputs, offering token-level guidance without requiring differentiable reasoning systems. This paradigm bridges the gap between symbolic reasoning and LLM fine-tuning, enabling precise alignment with domain-specific constraints while addressing key limitations of traditional reward signals. Via extensive evaluations, we show that our RLSF-based fine-tuning of LLMs outperforms traditional approaches on five different applications (that have some associated logical or domain constraints), namely, program synthesis from natural language pseudo-code to programming language, three chemistry tasks, and solving the Game of 24. A key takeaway is that fine-tuning via RLSF enables relatively smaller LLMs to significantly outperform closed-source models that are orders of magnitude larger.

cs.CL↗

arXiv2025

Abstractions-of-Thought: Intermediate Representations for LLM Reasoning in Hardware Design

Large language models (LLMs) have achieved impressive proficiency on logic and programming tasks, often rivaling expert-level performance. However, generating functionally correct hardware description language (HDL) code from natural language specifications remains challenging, primarily in data-scarce domains. Therefore, we present Abstractions-of-Thought (AoT) - a training-free, inference-only prompting framework to mitigate misinterpretations and reasoning pitfalls of LLMs through a series of task-based abstractions within the prompting procedure, assisting in the transition from high-level to low-level representations of hardware. Furthermore, AoT consists of the following stages: (1) an LLM-based classification of hardware design patterns, (2) a structured intermediate representation (IR) to separate functional decomposition from code syntax, and (3) a line-by-line pseudocode solution enabling a more direct mapping to the final Verilog implementation. Experimental results on the VerilogEval benchmark depict that AoT demonstrates improvements in functionality when applied to large non-reasoning models (such as GPT-4o, outperforming all baseline techniques (including 1-shot, Chain-of-Thought, and Tree-of-Thought) while significantly reducing the generated tokens by 1.8-5.2x compared to popular Tree-of-Thought prompting.

cs.PL↗

arXiv2024

CoTran: An LLM-based Code Translator using Reinforcement Learning with Feedback from Compiler and Symbolic Execution

In this paper, we present an LLM-based code translation method and an associated tool called CoTran, that translates whole-programs from one high-level programming language to another. Existing LLM-based code translation methods lack training to ensure that the translated code reliably compiles or bears substantial functional equivalence to the input code. In our work, we fine-tune an LLM using reinforcement learning, incorporating compiler feedback, and symbolic execution (symexec)-based testing feedback to assess functional equivalence between the input and output programs. The idea is to guide an LLM during fine-tuning, via compiler and symexec-based testing feedback, by letting it know how far it is from producing perfect translations. We conduct extensive experiments comparing CoTran with 14 other code translation tools, including human-written transpilers, LLM-based translation tools, and ChatGPT. Using a benchmark of over \num{57000} code pairs in Java and Python, we demonstrate that CoTran outperforms the other tools on relevant metrics such as compilation accuracy (CompAcc) and functional equivalence accuracy (FEqAcc). For example, in Python-to-Java translation, CoTran achieves 48.68% FEqAcc and 76.98% CompAcc, whereas the nearest competing tool (PLBART-base) gets 38.26% and 75.77% respectively. Additionally, CoTran, built on top of CodeT5, improves FEqAcc by +14.89% and CompAcc by +8.14% for Python-to-Java (resp., +12.94% and +4.30% for Java-to-Python).

cs.PL↗

arXiv2024

Layered and Staged Monte Carlo Tree Search for SMT Strategy Synthesis

Modern SMT solvers, such as Z3, offer user-controllable strategies, enabling users to tailor solving strategies for their unique set of instances, thus dramatically enhancing solver performance for their use case. However, this approach of strategy customization presents a significant challenge: handcrafting an optimized strategy for a class of SMT instances remains a complex and demanding task for both solver developers and users alike. In this paper, we address this problem of automatic SMT strategy synthesis via a novel Monte Carlo Tree Search (MCTS) based method. Our method treats strategy synthesis as a sequential decision-making process, whose search tree corresponds to the strategy space, and employs MCTS to navigate this vast search space. The key innovations that enable our method to identify effective strategies, while keeping costs low, are the ideas of layered and staged MCTS search. These novel heuristics allow for a deeper and more efficient exploration of the strategy space, enabling us to synthesize more effective strategies than the default ones in state-of-the-art (SOTA) SMT solvers. We implement our method, dubbed Z3alpha, as part of the Z3 SMT solver. Through extensive evaluations across six important SMT logics, Z3alpha demonstrates superior performance compared to the SOTA synthesis tool FastSMT, the default Z3 solver, and the CVC5 solver on most benchmarks. Remarkably, on a challenging QF_BV benchmark set, Z3alpha solves 42.7% more instances than the default strategy in the Z3 SMT solver.

cs.AI↗

arXiv2024

A Reinforcement Learning based Reset Policy for CDCL SAT Solvers

Restart policy is an important technique used in modern Conflict-Driven Clause Learning (CDCL) solvers, wherein some parts of the solver state are erased at certain intervals during the run of the solver. In most solvers, variable activities are preserved across restart boundaries, resulting in solvers continuing to search parts of the assignment tree that are not far from the one immediately prior to a restart. To enable the solver to search possibly "distant" parts of the assignment tree, we study the effect of resets, a variant of restarts which not only erases the assignment trail, but also randomizes the activity scores of the variables of the input formula after reset, thus potentially enabling a better global exploration of the search space. In this paper, we model the problem of whether to trigger reset as a multi-armed bandit (MAB) problem, and propose two reinforcement learning (RL) based adaptive reset policies using the Upper Confidence Bound (UCB) and Thompson sampling algorithms. These two algorithms balance the exploration-exploitation tradeoff by adaptively choosing arms (reset vs. no reset) based on their estimated rewards during the solver's run. We implement our reset policies in four baseline SOTA CDCL solvers and compare the baselines against the reset versions on Satcoin benchmarks and SAT Competition instances. Our results show that RL-based reset versions outperform the corresponding baseline solvers on both Satcoin and the SAT competition instances, suggesting that our RL policy helps to dynamically and profitably adapt the reset frequency for any given input instance. We also introduce the concept of a partial reset, where at least a constant number of variable activities are retained across reset boundaries. Building on previous results, we show that there is an exponential separation between O(1) vs. $Ω(n)$-length partial resets.

cs.LO↗

arXiv2024

BertRLFuzzer: A BERT and Reinforcement Learning Based Fuzzer

We present a novel tool BertRLFuzzer, a BERT and Reinforcement Learning (RL) based fuzzer aimed at finding security vulnerabilities for Web applications. BertRLFuzzer works as follows: given a set of seed inputs, the fuzzer performs grammar-adhering and attack-provoking mutation operations on them to generate candidate attack vectors. The key insight of BertRLFuzzer is the use of RL with a BERT model as an agent to guide the fuzzer to efficiently learn grammar-adhering and attack-provoking mutation operators. In order to establish the efficacy of BertRLFuzzer we compare it against a total of 13 black box and white box fuzzers over a benchmark of 9 victim websites with over 16K LOC. We observed a significant improvement relative to the nearest competing tool in terms of time to first attack (54% less), new vulnerabilities found (17 new vulnerabilities), and attack rate (4.4% more attack vectors generated).

cs.SE↗

arXiv2023

CGDTest: A Constrained Gradient Descent Algorithm for Testing Neural Networks

In this paper, we propose a new Deep Neural Network (DNN) testing algorithm called the Constrained Gradient Descent (CGD) method, and an implementation we call CGDTest aimed at exposing security and robustness issues such as adversarial robustness and bias in DNNs. Our CGD algorithm is a gradient-descent (GD) method, with the twist that the user can also specify logical properties that characterize the kinds of inputs that the user may want. This functionality sets CGDTest apart from other similar DNN testing tools since it allows users to specify logical constraints to test DNNs not only for $\ell_p$ ball-based adversarial robustness but, more importantly, includes richer properties such as disguised and flow adversarial constraints, as well as adversarial robustness in the NLP domain. We showcase the utility and power of CGDTest via extensive experimentation in the context of vision and NLP domains, comparing against 32 state-of-the-art methods over these diverse domains. Our results indicate that CGDTest outperforms state-of-the-art testing tools for $\ell_p$ ball-based adversarial robustness, and is significantly superior in testing for other adversarial robustness, with improvements in PAR2 scores of over 1500% in some cases over the next best tool. Our evaluation shows that our CGD method outperforms competing methods we compared against in terms of expressibility (i.e., a rich constraint language and concomitant tool support to express a wide variety of properties), scalability (i.e., can be applied to very large real-world models with up to 138 million parameters), and generality (i.e., can be used to test a plethora of model architectures).

cs.LG↗

arXiv2020

An Augmented Translation Technique for low Resource language pair: Sanskrit to Hindi translation

Neural Machine Translation (NMT) is an ongoing technique for Machine Translation (MT) using enormous artificial neural network. It has exhibited promising outcomes and has shown incredible potential in solving challenging machine translation exercises. One such exercise is the best approach to furnish great MT to language sets with a little preparing information. In this work, Zero Shot Translation (ZST) is inspected for a low resource language pair. By working on high resource language pairs for which benchmarks are available, namely Spanish to Portuguese, and training on data sets (Spanish-English and English-Portuguese) we prepare a state of proof for ZST system that gives appropriate results on the available data. Subsequently the same architecture is tested for Sanskrit to Hindi translation for which data is sparse, by training the model on English-Hindi and Sanskrit-English language pairs. In order to prepare and decipher with ZST system, we broaden the preparation and interpretation pipelines of NMT seq2seq model in tensorflow, incorporating ZST features. Dimensionality reduction of word embedding is performed to reduce the memory usage for data storage and to achieve a faster training and translation cycles. In this work existing helpful technology has been utilized in an imaginative manner to execute our NLP issue of Sanskrit to Hindi translation. A Sanskrit-Hindi parallel corpus of 300 is constructed for testing. The data required for the construction of parallel corpus has been taken from the telecasted news, published on Department of Public Information, state government of Madhya Pradesh, India website.

cs.CL↗

arXiv2017

Common Representation Learning Using Step-based Correlation Multi-Modal CNN

Deep learning techniques have been successfully used in learning a common representation for multi-view data, wherein the different modalities are projected onto a common subspace. In a broader perspective, the techniques used to investigate common representation learning falls under the categories of canonical correlation-based approaches and autoencoder based approaches. In this paper, we investigate the performance of deep autoencoder based methods on multi-view data. We propose a novel step-based correlation multi-modal CNN (CorrMCNN) which reconstructs one view of the data given the other while increasing the interaction between the representations at each hidden layer or every intermediate step. Finally, we evaluate the performance of the proposed model on two benchmark datasets - MNIST and XRMB. Through extensive experiments, we find that the proposed model achieves better performance than the current state-of-the-art techniques on joint common representation learning and transfer learning tasks.

cs.CV↗