arXiv subjects

Amrit Singh Bedi

Publications and source records attributed to Amrit Singh Bedi.

At least 37 records · Page 2Linked to original sources

arXiv2026

BalancedDPO: Adaptive Multi-Metric Alignment

Diffusion models have achieved remarkable progress in text-to-image generation, yet aligning them with human preference remains challenging due to the presence of multiple, sometimes conflicting, evaluation metrics (e.g., semantic consistency, aesthetics, and human preference scores). Existing alignment methods typically optimize for a single metric or rely on scalarized reward aggregation, which can bias the model toward specific evaluation criteria. To address this challenge, we propose BalancedDPO, a framework that achieves multi-metric preference alignment within the Direct Preference Optimization (DPO) paradigm. Unlike prior DPO variants that rely on a single metric, BalancedDPO introduces a majority-vote consensus over multiple preference scorers and integrates it directly into the DPO training loop with dynamic reference model updates. This consensus-based formulation avoids reward-scale conflicts and ensures more stable gradient directions across heterogeneous metrics. Experiments on Pick-a-Pic, PartiPrompt, and HPD datasets demonstrate that BalancedDPO consistently improves preference win rates over the baselines across Stable Diffusion 1.5, Stable Diffusion 2.1 and SDXL backbones. Comprehensive ablations further validate the benefits of majority-vote aggregation and dynamic reference updating, highlighting the method's robustness and generalizability across diverse alignment settings.

cs.CV↗

arXiv2026

Code Comprehension then Auditing for Unsupervised LLM Evaluation

Large Language Models (LLMs) for unsupervised code correctness evaluation have recently gained attention because they can judge if code runs as intended without requiring reference implementations or unit tests, which may be unavailable, sparse, or unreliable. However, most prior approaches condition LLM evaluators directly on the full code implementation, forcing the model to jointly infer program behavior and evaluate correctness in a single step. This entanglement leads to misinterpretations of code behavior and unreliable judgments. To mitigate this issue, we introduce CoCoA, an unsupervised Code Comprehension then Auditing framework that first comprehends functionality to generate a natural-language explanation. Then it evaluates task alignment based on this explanation. By sequentially sampling comprehension before evaluation, CoCoA improves the quality of inferred program behavior and enables the evaluator to focus on behavioral alignment rather than raw implementation details. Across multiple datasets, programming languages, and models, CoCoA achieves up to $68\%$ increased F1 score and up to $20\%$ increased accuracy over the best-performing baselines.

cs.AI↗

arXiv2026

On The Sample Complexity Bounds In Bilevel Reinforcement Learning

Bilevel reinforcement learning (BRL) has emerged as a powerful framework for aligning generative models, yet its theoretical foundations, especially sample complexity bounds, remain underexplored. In this work, we present the first sample complexity bound for BRL, establishing a rate of $\mathcal{O}(ε^{-3})$ in continuous state-action spaces. Traditional MDP analysis techniques do not extend to BRL due to its nested structure and non-convex lower-level problems. We overcome these challenges by leveraging the Polyak-Łojasiewicz (PL) condition and the MDP structure to obtain closed-form gradients, enabling tight sample complexity analysis. Our analysis also extends to general bi-level optimization settings with non-convex lower levels, where we achieve state-of-the-art sample complexity results of $\mathcal{O}(ε^{-3})$ improving upon existing bounds of $\mathcal{O}(ε^{-6})$. Additionally, we address the computational bottleneck of hypergradient estimation by proposing a fully first-order, Hessian-free algorithm suitable for large-scale problems.

cs.LG↗

arXiv2026

Safety Recovery in Reasoning Models Is Only a Few Early Steering Steps Away

Reinforcement learning (RL) based post-training for explicit chain-of-thought (e.g., GRPO) improves the reasoning ability of multimodal large-scale reasoning models (MLRMs). But recent evidence shows that it can simultaneously degrade safety alignment and increase jailbreak success rates. We propose SafeThink, a lightweight inference-time defense that treats safety recovery as a satisficing constraint rather than a maximization objective. SafeThink monitors the evolving reasoning trace with a safety reward model and conditionally injects an optimized short corrective prefix ("Wait, think safely") only when the safety threshold is violated. In our evaluations across six open-source MLRMs and four jailbreak benchmarks (JailbreakV-28K, Hades, FigStep, and MM-SafetyBench), SafeThink reduces attack success rates by 30-60% (e.g., LlamaV-o1: 63.33% to 5.74% on JailbreakV-28K, R1-Onevision: 69.07% to 5.65% on Hades) while preserving reasoning performance (MathVista accuracy: 65.20% to 65.00%). A key empirical finding from our experiments is that safety recovery is often only a few steering steps away: intervening in the first 1-3 reasoning steps typically suffices to redirect the full generation toward safe completions.

cs.CL↗

arXiv2025

PROPS: Progressively Private Self-alignment of Large Language Models

Alignment is a key step in developing Large Language Models (LLMs) using human feedback to ensure adherence to human values and societal norms. Dependence on human feedback raises privacy concerns about how much a labeler's preferences may reveal about their personal values, beliefs, and personality traits. Existing approaches, such as Differentially Private SGD (DP-SGD), provide rigorous privacy guarantees by privatizing gradients during fine-tuning and alignment but can provide more privacy than necessary as human preferences are tied only to labels of (prompt, response) pairs and can degrade model utility. This work focuses on LLM alignment with preference-level privacy, which preserves the privacy of preference labels provided by humans. We propose PROPS (PROgressively Private Self-alignment), a multi-stage privacy preserving alignment framework where privately aligned models in previous stages can serve as labelers for supplementing training data in the subsequent stages of alignment. We present theoretical guarantees for PROPS as well as comprehensive validation using multiple models (Pythia and GPT) and datasets (AlpacaEval, Anthropic HH-RLHF, truthy-dpo-v0.1) to demonstrate the utility of PROPS over existing methods while still providing high privacy. For the same privacy budget, alignment via PROPS can achieve up to 3x higher win-rates compared to DP-SGD, and 2.5x higher win-rates compared to Randomized Response (RR) based alignment.

cs.LG↗

arXiv2025

Generative Modeling with Continuous Flows: Sample Complexity of Flow Matching

Flow matching has recently emerged as a promising alternative to diffusion-based generative models, offering faster sampling and simpler training by learning continuous flows governed by ordinary differential equations. Despite growing empirical success, the theoretical understanding of flow matching remains limited, particularly in terms of sample complexity results. In this work, we provide the first analysis of the sample complexity for flow-matching based generative models without assuming access to the empirical risk minimizer (ERM) of the loss function for estimating the velocity field. Under standard assumptions on the loss function for velocity field estimation and boundedness of the data distribution, we show that a sufficiently expressive neural network can learn a velocity field such that with $\mathcal{O}(ε^{-4})$ samples, such that the Wasserstein-2 distance between the learned and the true distribution is less than $\mathcal{O}(ε)$. The key technical idea is to decompose the velocity field estimation error into neural-network approximation error, statistical error due to the finite sample size, and optimization error due to the finite number of optimization steps for estimating the velocity field. Each of these terms are then handled via techniques that may be of independent interest.

cs.LG↗

arXiv2025

SafeR-CLIP: Mitigating NSFW Content in Vision-Language Models While Preserving Pre-Trained Knowledge

Improving the safety of vision-language models like CLIP via fine-tuning often comes at a steep price, causing significant drops in their generalization performance. We find this trade-off stems from rigid alignment strategies that force unsafe concepts toward single, predefined safe targets, disrupting the model's learned semantic structure. To address this, we propose a proximity-aware approach: redirecting unsafe concepts to their semantically closest safe alternatives to minimize representational change. We introduce SaFeR-CLIP, a fine-tuning framework that applies this principle of minimal intervention. SaFeR-CLIP successfully reconciles safety and performance, recovering up to 8.0% in zero-shot accuracy over prior methods while maintaining robust safety. To support more rigorous evaluation, we also contribute NSFW-Caps, a new benchmark of 1,000 highly-aligned pairs for testing safety under distributional shift. Our work shows that respecting the geometry of pretrained representations is key to achieving safety without sacrificing performance.

cs.CV↗

arXiv2025

Generative AI for Biosciences: Emerging Threats and Roadmap to Biosecurity

The rapid adoption of generative artificial intelligence (GenAI) in the biosciences is transforming biotechnology, medicine, and synthetic biology. Yet this advancement is intrinsically linked to new vulnerabilities, as GenAI lowers the barrier to misuse and introduces novel biosecurity threats, such as generating synthetic viral proteins or toxins. These dual-use risks are often overlooked, as existing safety guardrails remain fragile and can be circumvented through deceptive prompts or jailbreak techniques. In this Perspective, we first outline the current state of GenAI in the biosciences and emerging threat vectors ranging from jailbreak attacks and privacy risks to the dual-use challenges posed by autonomous AI agents. We then examine urgent gaps in regulation and oversight, drawing on insights from 130 expert interviews across academia, government, industry, and policy. A large majority ($\approx 76$\%) expressed concern over AI misuse in biology, and 74\% called for the development of new governance frameworks. Finally, we explore technical pathways to mitigation, advocating a multi-layered approach to GenAI safety. These defenses include rigorous data filtering, alignment with ethical principles during development, and real-time monitoring to block harmful requests. Together, these strategies provide a blueprint for embedding security throughout the GenAI lifecycle. As GenAI becomes integrated into the biosciences, safeguarding this frontier requires an immediate commitment to both adaptive governance and secure-by-design technologies.

cs.CR↗

arXiv2025

On the Global Optimality of Policy Gradient Methods in General Utility Reinforcement Learning

Reinforcement learning with general utilities (RLGU) offers a unifying framework to capture several problems beyond standard expected returns, including imitation learning, pure exploration, and safe RL. Despite recent fundamental advances in the theoretical analysis of policy gradient (PG) methods for standard RL and recent efforts in RLGU, the understanding of these PG algorithms and their scope of application in RLGU still remain limited. In this work, we establish global optimality guarantees of PG methods for RLGU in which the objective is a general concave utility function of the state-action occupancy measure. In the tabular setting, we provide global optimality results using a new proof technique building on recent theoretical developments on the convergence of PG methods for standard RL using gradient domination. Our proof technique opens avenues for analyzing policy parameterizations beyond the direct policy parameterization for RLGU. In addition, we provide global optimality results for large state-action space settings beyond prior work which has mostly focused on the tabular setting. In this large scale setting, we adapt PG methods by approximating occupancy measures within a function approximation class using maximum likelihood estimation. Our sample complexity only scales with the dimension induced by our approximation class instead of the size of the state-action space.

cs.LG↗

arXiv2025

Does Thinking More always Help? Mirage of Test-Time Scaling in Reasoning Models

Recent trends in test-time scaling for reasoning models (e.g., OpenAI o1, DeepSeek R1) have led to a popular belief that extending thinking traces using prompts like "Wait" or "Let me rethink" can improve performance. This raises a natural question: Does thinking more at test-time truly lead to better reasoning? To answer this question, we perform a detailed empirical study across models and benchmarks, which reveals a consistent pattern of initial performance improvements from additional thinking followed by a decline, due to "overthinking". To understand this non-monotonic trend, we consider a simple probabilistic model, which reveals that additional thinking increases output variance-creating an illusion of improved reasoning while ultimately undermining precision. Thus, observed gains from "more thinking" are not true indicators of improved reasoning, but artifacts stemming from the connection between model uncertainty and evaluation metric. This suggests that test-time scaling through extended thinking is not an effective way to utilize the inference thinking budget. Recognizing these limitations, we introduce an alternative test-time scaling approach, parallel thinking, inspired by Best-of-N sampling. Our method generates multiple independent reasoning paths within the same inference budget and selects the most consistent response via majority vote, achieving up to 20% higher accuracy compared to extended thinking. This provides a simple yet effective mechanism for test-time scaling of reasoning models.

cs.AI↗

arXiv2025

Test-Time Scaling in Diffusion LLMs via Hidden Semi-Autoregressive Experts

Diffusion-based large language models (dLLMs) are trained flexibly to model extreme dependence in the data distribution; however, how to best utilize this information at inference time remains an open problem. In this work, we uncover an interesting property of these models: dLLMs trained on textual data implicitly learn a mixture of semi-autoregressive experts, where different generation orders reveal different specialized behaviors. We show that committing to any single, fixed inference time schedule, a common practice, collapses performance by failing to leverage this latent ensemble. To address this, we introduce HEX (Hidden semiautoregressive EXperts for test-time scaling), a training-free inference method that ensembles across heterogeneous block schedules. By doing a majority vote over diverse block-sized generation paths, HEX robustly avoids failure modes associated with any single fixed schedule. On reasoning benchmarks such as GSM8K, it boosts accuracy by up to 3.56X (from 24.72% to 88.10%), outperforming top-K margin inference and specialized fine-tuned methods like GRPO, without additional training. HEX even yields significant gains on MATH benchmark from 16.40% to 40.00%, scientific reasoning on ARC-C from 54.18% to 87.80%, and TruthfulQA from 28.36% to 57.46%. Our results establish a new paradigm for test-time scaling in diffusion-based LLMs (dLLMs), revealing that the sequence in which masking is performed plays a critical role in determining performance during inference.

cs.LG↗

arXiv2025

MIRA: Towards Mitigating Reward Hacking in Inference-Time Alignment of T2I Diffusion Models

Diffusion models excel at generating images conditioned on text prompts, but the resulting images often do not satisfy user-specific criteria measured by scalar rewards such as Aesthetic Scores. This alignment typically requires fine-tuning, which is computationally demanding. Recently, inference-time alignment via noise optimization has emerged as an efficient alternative, modifying initial input noise to steer the diffusion denoising process towards generating high-reward images. However, this approach suffers from reward hacking, where the model produces images that score highly, yet deviate significantly from the original prompt. We show that noise-space regularization is insufficient and that preventing reward hacking requires an explicit image-space constraint. To this end, we propose MIRA (MItigating Reward hAcking), a training-free, inference-time alignment method. MIRA introduces an image-space, score-based KL surrogate that regularizes the sampling trajectory with a frozen backbone, constraining the output distribution so reward can increase without off-distribution drift (reward hacking). We derive a tractable approximation to KL using diffusion scores. Across SDv1.5 and SDXL, multiple rewards (Aesthetic, HPSv2, PickScore), and public datasets (e.g., Animal-Animal, HPDv2), MIRA achieves >60\% win rate vs. strong baselines while preserving prompt adherence; mechanism plots show reward gains with near-zero drift, whereas DNO drifts as compute increases. We further introduce MIRA-DPO, mapping preference optimization to inference time with a frozen backbone, extending MIRA to non-differentiable rewards without fine-tuning.

cs.LG↗

arXiv2025

Uncertainty-Aware Answer Selection for Improved Reasoning in Multi-LLM Systems

Large Language Models (LLMs) have demonstrated exceptional capabilities, yet selecting the most reliable response from multiple LLMs remains a challenge, particularly in resource-constrained settings. Existing approaches often depend on costly external verifiers, human evaluators, or self-consistency techniques that require multiple samples from a single model. While multi-LLM systems produce more diverse responses than single models and thus have greater potential, they often underperform compared to single LLM self-consistency. We propose a principled, novel and computationally efficient method to select the best response from multiple different LLMs using a calibrated log-likelihood score, implicitly leveraging the inherent knowledge and confidence of these models. Our method demonstrates improvements of approx. 4%, 3%, and 5% across both debate (multi-round LLM discussions) and non-debate (Best-of-N with multiple LLMs) settings on GSM8K, MMLU (6 subsets), and ARC datasets respectively.

cs.CL↗

arXiv2025

EfficientEQA: An Efficient Approach to Open-Vocabulary Embodied Question Answering

Embodied Question Answering (EQA) is an essential yet challenging task for robot assistants. Large vision-language models (VLMs) have shown promise for EQA, but existing approaches either treat it as static video question answering without active exploration or restrict answers to a closed set of choices. These limitations hinder real-world applicability, where a robot must explore efficiently and provide accurate answers in open-vocabulary settings. To overcome these challenges, we introduce EfficientEQA, a novel framework that couples efficient exploration with free-form answer generation. EfficientEQA features three key innovations: (1) Semantic-Value-Weighted Frontier Exploration (SFE) with Verbalized Confidence (VC) from a black-box VLM to prioritize semantically important areas to explore, enabling the agent to gather relevant information faster; (2) a BLIP relevancy-based mechanism to stop adaptively by flagging highly relevant observations as outliers to indicate whether the agent has collected enough information; and (3) a Retrieval-Augmented Generation (RAG) method for the VLM to answer accurately based on pertinent images from the agent's observation history without relying on predefined choices. Our experimental results show that EfficientEQA achieves over 15% higher answer accuracy and requires over 20% fewer exploration steps than state-of-the-art methods. Our code is available at: https://github.com/chengkaiAcademyCity/EfficientEQA

cs.RO↗

arXiv2025

Leveraging Pre-Trained Visual Models for AI-Generated Video Detection

Recent advances in Generative AI (GenAI) have led to significant improvements in the quality of generated visual content. As AI-generated visual content becomes increasingly indistinguishable from real content, the challenge of detecting the generated content becomes critical in combating misinformation, ensuring privacy, and preventing security threats. Although there has been substantial progress in detecting AI-generated images, current methods for video detection are largely focused on deepfakes, which primarily involve human faces. However, the field of video generation has advanced beyond DeepFakes, creating an urgent need for methods capable of detecting AI-generated videos with generic content. To address this gap, we propose a novel approach that leverages pre-trained visual models to distinguish between real and generated videos. The features extracted from these pre-trained models, which have been trained on extensive real visual content, contain inherent signals that can help distinguish real from generated videos. Using these extracted features, we achieve high detection performance without requiring additional model training, and we further improve performance by training a simple linear classification layer on top of the extracted features. We validated our method on a dataset we compiled (VID-AID), which includes around 10,000 AI-generated videos produced by 9 different text-to-video models, along with 4,000 real videos, totaling over 7 hours of video content. Our evaluation shows that our approach achieves high detection accuracy, above 90% on average, underscoring its effectiveness. Upon acceptance, we plan to publicly release the code, the pre-trained models, and our dataset to support ongoing research in this critical area.

cs.CV↗

arXiv2025

On the Role of Feedback in Test-Time Scaling of Agentic AI Workflows

Agentic AI workflows (systems that autonomously plan and act) are becoming widespread, yet their task success rate on complex tasks remains low. A promising solution is inference-time alignment, which uses extra compute at test time to improve performance. Inference-time alignment relies on three components: sampling, evaluation, and feedback. While most prior work studies sampling and automatic evaluation, feedback remains underexplored. To study the role of feedback, we introduce Iterative Agent Decoding (IAD), a procedure that repeatedly inserts feedback extracted from different forms of critiques (reward models or AI-generated textual feedback) between decoding steps. Through IAD, we analyze feedback along four dimensions: (1) its role in the accuracy-compute trade-offs with limited inference budget, (2) quantifying the gains over diversity-only baselines such as best-of-N sampling, (3) effectiveness of composing feedback from reward models versus textual critique, and (4) robustness to noisy or low-quality feedback. Across Sketch2Code, Text2SQL, Intercode, and WebShop, we show that IAD with proper integration of high fidelity feedback leads to consistent gains up to 10 percent absolute performance improvement over various baselines such as best-of-N. Our findings underscore feedback as a crucial knob for inference-time alignment of agentic AI workflows with limited inference budget.

cs.CL↗

arXiv2025

LIAR: Leveraging Inference Time Alignment (Best-of-N) to Jailbreak LLMs in Seconds

Jailbreak attacks expose vulnerabilities in safety-aligned LLMs by eliciting harmful outputs through carefully crafted prompts. Existing methods rely on discrete optimization or trained adversarial generators, but are slow, compute-intensive, and often impractical. We argue that these inefficiencies stem from a mischaracterization of the problem. Instead, we frame jailbreaks as inference-time misalignment and introduce LIAR (Leveraging Inference-time misAlignment to jailbReak), a fast, black-box, best-of-$N$ sampling attack requiring no training. LIAR matches state-of-the-art success rates while reducing perplexity by $10\times$ and Time-to-Attack from hours to seconds. We also introduce a theoretical "safety net against jailbreaks" metric to quantify safety alignment strength and derive suboptimality bounds. Our work offers a simple yet effective tool for evaluating LLM robustness and advancing alignment research.

cs.CL↗

arXiv2025

Immune: Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment

With the widespread deployment of Multimodal Large Language Models (MLLMs) for visual-reasoning tasks, improving their safety has become crucial. Recent research indicates that despite training-time safety alignment, these models remain vulnerable to jailbreak attacks. In this work, we first highlight an important safety gap to describe that alignment achieved solely through safety training may be insufficient against jailbreak attacks. To address this vulnerability, we propose Immune, an inference-time defense framework that leverages a safe reward model through controlled decoding to defend against jailbreak attacks. Additionally, we provide a mathematical characterization of Immune, offering insights on why it improves safety against jailbreaks. Extensive evaluations on diverse jailbreak benchmarks using recent MLLMs reveal that Immune effectively enhances model safety while preserving the model's original capabilities. For instance, against text-based jailbreak attacks on LLaVA-1.6, Immune reduces the attack success rate by 57.82% and 16.78% compared to the base MLLM and state-of-the-art defense strategy, respectively.

cs.CR↗