arXiv subjects

Rishabh Bhardwaj

Publications and source records attributed to Rishabh Bhardwaj.

At least 19 recordsLinked to original sources

arXiv2026

Post-Reasoning: Improving the Performance of Non-Thinking Models at No Cost

As the widespread adoption of Large Language Models (LLMs) accelerates, token consumption from intermediate reasoning traces increasingly contributes to inference latency and operational cost. Recent studies suggest that many real-world tasks require little to no explicit reasoning, with additional reasoning sometimes even degrading performance. In this work, we propose Post-Reasoning, a simple yet effective approach that improves instruction-tuned models by conditioning them to justify their answers after generating the final response. By design, it enables the final answer to be obtained without additional latency or token cost, while still improving performance through simple instruction augmentation. We evaluate Post-Reasoning across 117 model--benchmark settings spanning 13 open and proprietary models, 4 model families, and 9 diverse reasoning and knowledge-intensive benchmarks, including AMC, HMMT, GSM8K, GPQA, MMLU-Pro, and BIG-Bench Hard. Post-Reasoning improves performance in over 88.19% of evaluated settings, achieving a mean relative improvements of 17.37%. Furthermore, we propose supervised post-reason tuning, which further improves performance in over 91.11% of evaluated settings, and exceeds the prompt-based post-reasoning baseline by an average of 8.01%, demonstrating that post-reasoning can be effectively internalized through training. Ultimately, Post-Reasoning establishes a new performance ceiling for direct-answer capabilities.

cs.AI↗

arXiv2026

Learning Low-Energy Subspace Overlaps in Many-Body Systems with Measurement-Based and Coherent Quantum Strategies

Predicting the overlap of quantum states with specified low-energy subspaces is a key diagnostic for quantum many-body dynamics, with direct applications in state preparation, subspace-based algorithms, and the study of thermalization. We study the supervised prediction of subspace overlaps O_K between time-evolved states and K-dimensional low-energy eigenspaces of a 10-qubit Heisenberg spin chain following a local perturbation. We compare two quantum information extraction strategies: measurement-based learning, in which classical shadow features are processed by convolutional neural networks, and coherent quantum learning, in which quantum convolutional neural networks process the state directly. We further introduce physics-informed variants for both approaches, including Hamiltonian-aware shadows and QCNN gates aligned with the Heisenberg exchange structure. Across five dataset configurations spanning weak, moderate, and strong quench regimes, physics-informed QCNNs achieve stable performance, with mean test-set coefficients of determination R^2 = 0.753-0.846. Shadow-based methods show stronger regime dependence: they outperform QCNNs in the moderate-quench regime, reaching R^2 = 0.886, but underperform in weak and strong quenches at default shot budgets, where the best shadow results are R^2 = 0.615 and 0.672, respectively. Hardware validation on Quantinuum and IBM noise models shows that arbitrary state preparation is the dominant limitation, requiring approximately 2,044 two-qubit gates and causing near-complete depolarization before inference. These results identify a regime-dependent tradeoff between measurement-based and coherent quantum learning, with shadow methods excelling when the target remains locally accessible and physics-informed QCNNs providing more robust performance across dynamical regimes.

quant-ph↗

arXiv2026

Fault tolerant computation of the static structure factor and finite size effects

Fault-tolerant quantum algorithms offer a promising pathway for estimating the ground-state energies of periodic materials that are beyond the practical reach of classical electronic-structure methods. A remaining challenge is finite-size mitigation: quantum algorithms evaluate a finite supercell or finite Brillouin-zone mesh, while materials properties are defined in the thermodynamic limit. In this work we develop a quantum post-processing strategy for the leading two-body finite-size correction. After one-body shell effects are reduced by twist averaging, the dominant residual error is controlled by long-wavelength density fluctuations, which are encoded in the small-momentum static structure factor $S(q)$. We formulate the corresponding operator in a Bloch-orbital basis, construct its block encoding through the density operator, and estimate its ground-state expectation value using an amplified Hadamard test. We also introduce adaptive global and local binary search procedures for identifying the infrared fitting window used to reconstruct the two-body finite size error correction. The resulting cost remains subleading relative to the main ground-state energy estimation routine: the structure-factor correction has leading $\tilde{O}(N_bN_k)^3$ dependence on the Bloch-orbital basis size, avoids the large plane-wave prefactor of full Hamiltonian simulation, and requires only $\tilde{O}(N_bN_k)$ logical qubits. This provides a fault-tolerant alternative to down-sampling, replacing repeated energy calculations on larger cells with targeted measurements of the infrared density correlations that control the finite-size effects.

quant-ph↗

arXiv2026

The Quantum Hamiltonian Analysis Toolkit: Lowering the Barrier to Quantum Computing with Hamiltonians

We present the Quantum Hamiltonian Analysis Toolkit (QHAT), a newly developed application that provides a user-friendly interface for studying Hamiltonians and performing Hamiltonian simulation on fault-tolerant quantum computers. QHAT enables the generation and analysis of Hamiltonians through a powerful and feature-rich application, driven by simple inputs designed to reflect user needs rather than algorithmic details, so that productive research on your application of interest can be done without needing a deep understanding of quantum computing algorithms. QHAT enables a streamlined workflow to analyze Hamiltonians and Hamiltonian simulation, supporting multiple choices of algorithms and analyses. It supports Hamiltonians from multiple sources but can also generate Hamiltonians based on a simple description of the system, saving intermediate data files for re-use when generating related Hamiltonians. Deriving the parameters for quantum computing algorithms can be a challenge, so QHAT is built around user-facing concepts such as maximum allowable error, rather than being built around algorithmic details such as steps counts or order parameters. An emphasis on user-friendly interfaces and efficient analysis means that the barrier to entry is low while rapidly providing results useful for a broad scope of studies.

quant-ph↗

arXiv2026

Fault-tolerant simulation of the electronic structure using Projector Augmented-Waves and Bloch orbitals

Strongly correlated materials are a natural target for fault-tolerant quantum computers, but they require tools beyond those developed for molecules. Electronic wavefunctions vary rapidly near nuclei yet remain delocalized across many unit cells, and bulk properties must be converged systematically with respect to finite-size errors. To resolve such issues, we present the Bloch--UPAW framework that combines Bloch-orbital $k$-space structure with unitary projector-augmented-wave (UPAW) augmentation. The UPAW Hamiltonian, expressed directly in the Bloch basis, retains explicit control of Brillouin-zone sampling, and incorporates near-nuclear physics through strictly local on-site corrections. The construction is independent of the underlying one-particle representation, so it applies to both plane-wave and localized bases, and it handles supercells for symmetry-breaking phenomena more efficiently. We derive a linear-combination-of-unitaries decomposition and a block-encoding circuit suitable for qubitization; UPAW augmentation adds one ancilla qubit and no Toffoli gates at leading order relative to a Bloch-only block encoding. Asymptotically, the Toffoli cost scales as $\mathcal{O}(N_k^3)$ when refining the $k$-mesh and as $\mathcal{O}(N_a^{3.5})$ when enlarging the supercell, enabling convergence to be steered by the most favorable route for a given material. Resource estimates for bulk diamond show approximately an order-of-magnitude reduction in Toffoli count relative to prior work on periodic solids.

quant-ph↗

arXiv2026

OffTopicEval: When Large Language Models Enter the Wrong Chat, Almost Always!

Large Language Model (LLM) safety is one of the most pressing challenges for enabling wide-scale deployment. While most studies and global discussions focus on generic harms, such as models assisting users in harming themselves or others, enterprises face a more fundamental concern: whether LLM-based agents are safe for their intended use case. To address this, we introduce operational safety, defined as an LLM's ability to appropriately accept or refuse user queries when tasked with a specific purpose. We further propose OffTopicEval, an evaluation suite and benchmark for measuring operational safety both in general and within specific agentic use cases. Our evaluations on six model families comprising 20 open-weight LLMs reveal that while performance varies across models, all of them remain highly operationally unsafe. Even the strongest models - Qwen-3 (235B) with 77.77% and Mistral (24B) with 79.96% - fall far short of reliable operational safety, while GPT models plateau in the 62-73% range, Phi achieves only mid-level scores (48-70%), and Gemma and Llama-3 collapse to 39.53% and 23.84%, respectively. While operational safety is a core model alignment issue, to suppress these failures, we propose prompt-based steering methods: query grounding (Q-ground) and system-prompt grounding (P-ground), which substantially improve OOD refusal. Q-ground provides consistent gains of up to 23%, while P-ground delivers even larger boosts, raising Llama-3.3 (70B) by 41% and Qwen-3 (30B) by 27%. These results highlight both the urgent need for operational safety interventions and the promise of prompt-based steering as a first step toward more reliable LLM-based agents.

cs.AI↗

arXiv2025

Quantum Data Learning of Topological-to-Ferromagnetic Phase Transitions in the 2+1D Toric Code Loop Gas Model

Quantum data learning (QDL) provides a framework for extracting physical insights directly from quantum states, bypassing the need for any identification of the classical observable of the theory. A central challenge in many-body physics is that the identity of quantum phases, especially those with topological order, are often inaccessible through local observables or simple symmetry-breaking diagnostics. Here, we apply QDL techniques to the 2+1-dimensional toric-code loop-gas model in a magnetic field. Ground states are generated across multiple lattice sizes using a parametrized loop-gas circuit (PLGC) with a variational quantum-eigensolver (VQE) approach. We then train a quantum convolutional neural network (QCNN) across the full field-parameter range to perform phase classification and capture the overall phase structure. We also employ a physics-aware training protocol that excludes the near-critical region (0.2 <= x <= 0.4)) around (x_c = 0.25), the phase-transition point estimated by quantum Monte Carlo, reserving this window for testing to evaluate the ability of the model to learn the phase transition. In parallel, we implement an unsupervised quantum k-means method based on state overlaps, which partitions the dataset into two phases without prior labeling. Our supervised QDL approach recovers the phase structure and accurately locates the phase transition, in close agreement with previously reported values; the unsupervised QDL approach recovers the phase structure and locates the phase transition with a small offset as expected in finite volumes; both QDL methods outperform classical alternatives. These findings establish QDL as an effective framework for characterizing topological quantum matter, studying finite volume effects, and probing phase diagrams of higher-dimensional systems.

quant-ph↗

arXiv2025

Evaluating AI for Finance: Is AI Credible at Assessing Investment Risk?

We assess whether AI systems can credibly evaluate investment risk appetite-a task that must be thoroughly validated before automation. Our analysis was conducted on proprietary systems (GPT, Claude, Gemini) and open-weight models (LLaMA, DeepSeek, Mistral), using carefully curated user profiles that reflect real users with varying attributes such as country and gender. As a result, the models exhibit significant variance in score distributions when user attributes-such as country or gender-that should not influence risk computation are changed. For example, GPT-4o assigns higher risk scores to Nigerian and Indonesian profiles. While some models align closely with expected scores in the Low- and Mid-risk ranges, none maintain consistent scores across regions and demographics, thereby violating AI and finance regulations.

cs.CL↗

arXiv2025

Measuring and Enhancing Trustworthiness of LLMs in RAG through Grounded Attributions and Learning to Refuse

LLMs are an integral component of retrieval-augmented generation (RAG) systems. While many studies focus on evaluating the overall quality of end-to-end RAG systems, there is a gap in understanding the appropriateness of LLMs for the RAG task. To address this, we introduce Trust-Score, a holistic metric that evaluates the trustworthiness of LLMs within the RAG framework. Our results show that various prompting methods, such as in-context learning, fail to effectively adapt LLMs to the RAG task as measured by Trust-Score. Consequently, we propose Trust-Align, a method to align LLMs for improved Trust-Score performance. 26 out of 27 models aligned using Trust-Align substantially outperform competitive baselines on ASQA, QAMPARI, and ELI5. Specifically, in LLaMA-3-8b, Trust-Align outperforms FRONT on ASQA (up 12.56), QAMPARI (up 36.04), and ELI5 (up 17.69). Trust-Align also significantly enhances models' ability to correctly refuse and provide quality citations. We also demonstrate the effectiveness of Trust-Align across different open-weight models, including the LLaMA series (1b to 8b), Qwen-2.5 series (0.5b to 7b), and Phi3.5 (3.8b). We release our code at https://github.com/declare-lab/trust-align.

cs.CL↗

arXiv2025

MSTS: A Multimodal Safety Test Suite for Vision-Language Models

Vision-language models (VLMs), which process image and text inputs, are increasingly integrated into chat assistants and other consumer AI applications. Without proper safeguards, however, VLMs may give harmful advice (e.g. how to self-harm) or encourage unsafe behaviours (e.g. to consume drugs). Despite these clear hazards, little work so far has evaluated VLM safety and the novel risks created by multimodal inputs. To address this gap, we introduce MSTS, a Multimodal Safety Test Suite for VLMs. MSTS comprises 400 test prompts across 40 fine-grained hazard categories. Each test prompt consists of a text and an image that only in combination reveal their full unsafe meaning. With MSTS, we find clear safety issues in several open VLMs. We also find some VLMs to be safe by accident, meaning that they are safe because they fail to understand even simple test prompts. We translate MSTS into ten languages, showing non-English prompts to increase the rate of unsafe model responses. We also show models to be safer when tested with text only rather than multimodal prompts. Finally, we explore the automation of VLM safety assessments, finding even the best safety classifiers to be lacking.

cs.CL↗

arXiv2024

Libra-Leaderboard: Towards Responsible AI through a Balanced Leaderboard of Safety and Capability

To address this gap, we introduce Libra-Leaderboard, a comprehensive framework designed to rank LLMs through a balanced evaluation of performance and safety. Combining a dynamic leaderboard with an interactive LLM arena, Libra-Leaderboard encourages the joint optimization of capability and safety. Unlike traditional approaches that average performance and safety metrics, Libra-Leaderboard uses a distance-to-optimal-score method to calculate the overall rankings. This approach incentivizes models to achieve a balance rather than excelling in one dimension at the expense of some other ones. In the first release, Libra-Leaderboard evaluates 26 mainstream LLMs from 14 leading organizations, identifying critical safety challenges even in state-of-the-art models.

cs.CL↗

arXiv2024

On Unitarity of Bespoke Amplitudes

We use partial wave unitarity to constrain various bespoke four-point amplitudes. We start by constructing bespoke generalizations of the type I superstring amplitude, which we show satisfy dual resonance and have suitable high-energy limits. By analyzing the behavior of partial wave coefficients for highly massive states, we strictly rule out all bespoke amplitudes with asymptotically non-linear Regge trajectories and place constraints on the first few non-trivial parameters in asymptotically linear cases. Finally, we argue that while a large class of unitary bespoke amplitudes fails to satisfy Regge Sum Rules, there exists a smaller sub-class with a vanishing mass gap that is superpolynomially bounded.

hep-th↗

arXiv2024

Ferret: Faster and Effective Automated Red Teaming with Reward-Based Scoring Technique

In today's era, where large language models (LLMs) are integrated into numerous real-world applications, ensuring their safety and robustness is crucial for responsible AI usage. Automated red-teaming methods play a key role in this process by generating adversarial attacks to identify and mitigate potential vulnerabilities in these models. However, existing methods often struggle with slow performance, limited categorical diversity, and high resource demands. While Rainbow Teaming, a recent approach, addresses the diversity challenge by framing adversarial prompt generation as a quality-diversity search, it remains slow and requires a large fine-tuned mutator for optimal performance. To overcome these limitations, we propose Ferret, a novel approach that builds upon Rainbow Teaming by generating multiple adversarial prompt mutations per iteration and using a scoring function to rank and select the most effective adversarial prompt. We explore various scoring functions, including reward models, Llama Guard, and LLM-as-a-judge, to rank adversarial mutations based on their potential harm to improve the efficiency of the search for harmful mutations. Our results demonstrate that Ferret, utilizing a reward model as a scoring function, improves the overall attack success rate (ASR) to 95%, which is 46% higher than Rainbow Teaming. Additionally, Ferret reduces the time needed to achieve a 90% ASR by 15.2% compared to the baseline and generates adversarial prompts that are transferable i.e. effective on other LLMs of larger size. Our codes are available at https://github.com/declare-lab/ferret.

cs.CL↗

arXiv2024

WalledEval: A Comprehensive Safety Evaluation Toolkit for Large Language Models

WalledEval is a comprehensive AI safety testing toolkit designed to evaluate large language models (LLMs). It accommodates a diverse range of models, including both open-weight and API-based ones, and features over 35 safety benchmarks covering areas such as multilingual safety, exaggerated safety, and prompt injections. The framework supports both LLM and judge benchmarking and incorporates custom mutators to test safety against various text-style mutations, such as future tense and paraphrasing. Additionally, WalledEval introduces WalledGuard, a new, small, and performant content moderation tool, and two datasets: SGXSTest and HIXSTest, which serve as benchmarks for assessing the exaggerated safety of LLMs and judges in cultural contexts. We make WalledEval publicly available at https://github.com/walledai/walledeval.

cs.CL↗

arXiv2024

A double copy from twisted (co)homology at genus one

We study the twisted (co)homology of a family of genus-one integrals -- the so called Riemann-Wirtinger integrals. These integrals are closely related to one-loop string amplitudes in chiral splitting where one leaves the loop-momentum, modulus and all but one puncture un-integrated. While not actual one-loop string integrals, they share many properties and are simple enough that the associated twisted (co)homologies have been completely characterized [Goto2022,arXiv:2206.03177]. Using intersection numbers -- an inner product on the vector space of allowed differential forms -- we derive the Gauss-Manin connection for two bases of the twisted cohomology providing an independent check of [Mano&Watanabe2012]. We also use the intersection index -- an inner product on the vector space of allowed contours -- to derive a double-copy formula for the closed-string analogues of Riemann-Wirtinger integrals (one-dimensional integrals over the torus). Similar to the celebrated KLT formula between open- and closed-string tree-level amplitudes, these intersection indices form a genus-one KLT-like kernel defining bilinears in meromorphic Riemann-Wirtinger integrals that are equal to their complex counterparts.

hep-th↗

arXiv2024

DELLA-Merging: Reducing Interference in Model Merging through Magnitude-Based Sampling

With the proliferation of domain-specific models, model merging has emerged as a set of techniques that combine the capabilities of multiple models into one that can multitask without the cost of additional training. In this paper, we propose a new model merging technique, Drop and rEscaLe via sampLing with mAgnitude (DELLA-Merging), that employs a novel pruning technique, MAGPRUNE, which shows significant advantages over DARE and TIES. MAGPRUNE first ranks the parameters in order of their magnitude and assigns higher dropout probabilities (p) to parameters with lower ranks corresponding to lower magnitudes. To approximate the original embeddings, MAGPRUNE employs a rescaling operation on the parameters that survive the random dropping by 1/(1 - p). On three different expert models considered for merging (LM, Math, Code) and corresponding benchmark datasets (AlpacaEval, GSM8K, MBPP), DELLA shows an average improvement of 2.4 points over baseline methods employing delta parameter pruning (an improvement of 3.6 points over TIES, 1.2 points over DARE), and 11.1 points over the no-pruning baseline (TA). We release the source code at: https://github.com/declare-lab/della.

cs.CL↗

arXiv2024

Ruby Teaming: Improving Quality Diversity Search with Memory for Automated Red Teaming

We propose Ruby Teaming, a method that improves on Rainbow Teaming by including a memory cache as its third dimension. The memory dimension provides cues to the mutator to yield better-quality prompts, both in terms of attack success rate (ASR) and quality diversity. The prompt archive generated by Ruby Teaming has an ASR of 74%, which is 20% higher than the baseline. In terms of quality diversity, Ruby Teaming outperforms Rainbow Teaming by 6% and 3% on Shannon's Evenness Index (SEI) and Simpson's Diversity Index (SDI), respectively.

cs.CL↗

arXiv2024

HyperTTS: Parameter Efficient Adaptation in Text to Speech using Hypernetworks

Neural speech synthesis, or text-to-speech (TTS), aims to transform a signal from the text domain to the speech domain. While developing TTS architectures that train and test on the same set of speakers has seen significant improvements, out-of-domain speaker performance still faces enormous limitations. Domain adaptation on a new set of speakers can be achieved by fine-tuning the whole model for each new domain, thus making it parameter-inefficient. This problem can be solved by Adapters that provide a parameter-efficient alternative to domain adaptation. Although famous in NLP, speech synthesis has not seen much improvement from Adapters. In this work, we present HyperTTS, which comprises a small learnable network, "hypernetwork", that generates parameters of the Adapter blocks, allowing us to condition Adapters on speaker representations and making them dynamic. Extensive evaluations of two domain adaptation settings demonstrate its effectiveness in achieving state-of-the-art performance in the parameter-efficient regime. We also compare different variants of HyperTTS, comparing them with baselines in different studies. Promising results on the dynamic adaptation of adapter parameters using hypernetworks open up new avenues for domain-generic multi-speaker TTS systems. The audio samples and code are available at https://github.com/declare-lab/HyperTTS.

cs.CL↗