Research
On-device research index

arXiv research

A locally-built, LLM-digested index of recent arXiv papers in quant finance, geometry/topology, and statistical ML — keyword search served straight from SQLite on this machine.

168,695 papers · 148 categories

Trend · papers per month

9182635 · Oct 202519922001200920172026
48 results for Harmful Prompts

Diffusion LLMs can efficiently generate harmful prompts for adversarial testing.

problem Generating harmful prompts for adversarial testing is resource-intensive and costly.
method Transformed adversarial prompt optimization into an efficient inference task using pretrained Diffusion LLMs.
result Only a few conditional samples are required to generate harmful prompts with high reward.

Shorter adversarial prompts help protect LLMs from jailbreak attacks.

problem Protecting large language models from jailbreak attacks with long adversarial suffixes.
method Adversarial training on shorter adversarial suffixes to defend against longer adversarial suffixes.
result Aligning LLMs on shorter adversarial suffixes can effectively defend against jailbreak attacks with longer suffixes.

COCA refactors training data to identify and erase unsafe concepts in LLMs.

problem Identifying and erasing unsafe concepts in Large Language Models (LLMs) for safety alignment.
method Concept Concentration (COCA) refactors training data with an explicit reasoning process to identify and erase unsafe concepts.
result COCA significantly reduces both in-distribution and out-of-distribution jailbreak success rates while maintaining strong performance on regular tasks.

Fine-tuning harms in-context learning, but restricting updates to the value matrix improves zero-shot performance.

problem Fine-tuning harms in-context learning, reducing zero-shot performance on unseen tasks.
method Theoretical analysis of linear attention models, identifying conditions for degraded few-shot performance.
result Restricting updates to the value matrix improves zero-shot performance while preserving in-context learning.

VERA-V uses variational inference to discover vulnerabilities in multimodal vision-language models.

problem Existing methods for jailbreaking vision-language models are brittle, limited, and focus on single attacks.
method VERA-V recasts jailbreak discovery as learning a joint posterior distribution over text-image prompts, using variational inference and three complementary strategies.
result VERA-V consistently outperforms state-of-the-art baselines, achieving up to 53.75% higher attack success rate.

Proposes a new method to measure and avoid harm in machine learning decisions.

problem Measuring and avoiding harm in machine learning algorithms.
method Formal definition of harm and benefit using causal models, counterfactual objective functions.
result Demonstrates that standard machine learning methods can lead to harmful policies under distributional shifts.

New framework improves interpretability of trainable prompts.

problem Improving task-specific LLM performance with soft prompts remains a black-box method.
method Developed a theoretical framework for evaluating interpretability of trainable prompts, inspired new objective functions.
result Found a fundamental trade-off between interpretability and task performance in trainable prompts.

Paper introduces Latent-CLIP for efficient text-image comparison in latent space.

problem Efficiently compare text and images in latent space without costly decoding.
method Trains CLIP model in latent space, uses Latent-CLIP rewards for noise optimization, and guides generation away from harmful content.
result Latent-CLIP matches CLIP performance on text-image classification and harmful content detection.

This paper formalizes AI safety using hypothesis testing in GenAI.

problem Ensuring safety of generative AI tools that create realistic content.
method Formalization of computational safety through hypothesis testing and signal processing.
result Demonstrates how AI safety can be assessed quantitatively using mathematical frameworks.

Study on how attention in prompt-tuning affects large language models.

problem Limited theoretical understanding of prompt-tuning and attention in LLMs.
method Exploration of prompt-tuning for one-layer attention architectures, contextual mixture-models, and self-contained prompt-attention model.
result Softmax-prompt-attention is more expressive than self-attention and linear-prompt-attention under contextual data model.

InfoPrompt improves soft prompt tuning by maximizing mutual information, leading to better performance.

problem High sensitivity of prompt tuning to initial conditions and insufficient task-relevant information.
method Develops an information-theoretic framework to maximize mutual information between prompts and model parameters, using novel loss functions.
result InfoPrompt accelerates convergence and outperforms traditional methods.

Understanding optimal prompts for binary sequence predictors is challenging.

problem Finding good prompts for binary sequence predictors is difficult.
method Viewing prompting as finding the best conditioning sequence on a near-optimal sequence predictor, using empirical and statistical analysis.
result Optimal prompts can be better understood given the pretraining distribution, which is not usually available.

Training models to prefer certain responses can unintentionally shift probability to harmful ones.

problem Likelihood displacement in DPO models, leading to unintended unalignment.
method Characterized and mitigated likelihood displacement using CHES score.
result Training models to prefer certain responses can unintentionally shift probability mass to harmful responses.

Contrastive learning harms minority group representations, affecting downstream tasks.

problem Representation harm in contrastive learning, especially affecting minority groups.
method Causal mediation analysis and stochastic block model explanation.
result Representation harm in contrastive learning is partly responsible for allocation harm in downstream tasks.

Transformer models can approximate smooth functions with prompts, enhancing LLMs' dynamic capabilities.

problem Lack of theoretical framework for prompt engineering in transformer models.
method Formal framework demonstrating transformer models can approximate ββ-times differentiable functions with prompts.
result Transformer models can approximate ββ-times differentiable functions with arbitrary precision using appropriately structured prompts.

Single-head transformers with a single self-attention layer can approximate any sequence-to-sequence function and are efficient under certain conditions.

problem Statistical and computational limits of prompt tuning for transformer-based models.
method Investigation of single-head transformers with a single self-attention layer, proving universality and efficiency under SETH.
result Existence of almost-linear time prompt tuning inference algorithms under certain conditions.

TRIPLE efficiently optimizes prompts with a budget constraint.

problem Efficiently selecting good prompts from a pool of candidates.
method TRIPLE connects prompt optimization to best arm identification in MAB, leveraging BAI-FB tools.
result TRIPLE outperforms baselines on multiple tasks with limited budget constraints.

Reward collapse occurs when ranking-based reward models yield uniform rewards for different prompts.

problem Reward collapse in aligning large language models with human preferences.
method Introduced a prompt-aware optimization scheme to derive closed-form expressions for reward distributions.
result Our prompt-aware utility functions significantly alleviate reward collapse during training.

eva method improves RL models by adaptively creating prompts, boosting performance.

problem Fixed prompt distribution limits scalability of RL models post-training.
method Evolving Alignment via Asymmetric Self-Play (eva) approach.
result Significant performance boost on benchmarks, e.g. 51.6% to 60.1% win-rate.

The paper studies estimation rates for MoE models with a new prompt.

problem Estimating parameters in a softmax-contaminated MoE model.
method Analytic notion of distinguishability, minimax optimal estimation rates.
result Estimation rates are minimax optimal under distinguishability, but slower otherwise.

LanguaShrink compresses prompts using psycholinguistic principles to reduce costs.

problem Reducing computational cost and efficiency in large language model inference.
method Leverages psycholinguistic principles and the Ebbinghaus memory curve to compress prompts.
result Achieves up to 26 times compression while maintaining semantic similarity.

VERA uses variational inference to jailbreak LLMs without manual optimization.

problem Lack of principled objective for gradient-based optimization in jailbreaking LLMs.
method VERA casts black-box jailbreak prompting as a variational inference problem, training a small attacker LLM to approximate the target LLM's posterior over adversarial prompts.
result VERA achieves strong performance across various target LLMs, demonstrating the value of probabilistic inference for adversarial prompt generation.

Develops PromptShift-CRC for drift-aware conformal risk control in foundation models under prompt and domain shift.

problem Fixed calibration risk in foundation models due to prompt and domain shift.
method Embeds prompts and responses, measures drift, gives more weight to recent examples, and updates risk online.
result Develops method to control risk up to terms for distribution mismatch and weighted quantile uncertainty.

Gradient-based framework for optimizing text prompts in diffusion models.

problem Efficiently optimizing prompts in text-to-image diffusion models with large domain space and non-differentiable embeddings.
method Formulated as discrete optimization over language space, designed compact subspaces, and introduced shortcut text gradient.
result Empirically discovered prompts that enhance or destroy image faithfulness.

ZegOT uses optimal transport to zero-shot segment images with text prompts.

problem Zero-shot semantic segmentation with limited image-text alignment knowledge.
method ZegOT uses optimal transport to match multiple text prompts with frozen image embeddings.
result ZegOT achieves state-of-the-art performance in zero-shot semantic segmentation.

This work analyzes CoT prompting methods from a statistical estimation perspective.

problem Improving the effectiveness of LLMs in solving multi-step reasoning problems.
method Introducing a multi-step latent variable model to characterize CoT prompting from a statistical estimation viewpoint.
result The CoT estimator is equivalent to a Bayesian estimator when the pretraining dataset is large.

Transformers can emulate various algorithms by prompting, proving universality.

problem How to emulate algorithms using fixed-weight Transformers.
method Two modes of in-context algorithm emulation: task-specific and prompt-programmable. Constructing prompts that encode algorithm parameters into token representations.
result Fixed-weight Transformers can emulate a broad class of algorithms via prompts.

Prediction models can harm patients even when accurate, leading to self-fulfilling prophecies.

problem Prediction models can lead to harmful decisions that worsen patient outcomes.
method Formal characterization of harmful prediction models and analysis of their impact.
result Well-calibrated models are ineffective for decision-making as they do not change the data distribution.

OTSeg uses multi-prompt Sinkhorn attention to improve zero-shot semantic segmentation.

problem Leveraging pre-trained CLIP knowledge to align text embeddings with pixel embeddings.
method OTSeg employs Multi-Prompts Sinkhorn (MPS) and Multi-Prompts Sinkhorn Attention (MPSA) to enhance semantic feature matching.
result OTSeg achieves state-of-the-art performance in zero-shot semantic segmentation tasks.

Study analyzes convergence of parameter estimation in contaminated mixture of experts.

problem Challenges in learning from prompts in large-scale models.
method Convergence analysis, distinguishability condition, partial differential equations.
result Comprehensive convergence rates and minimax lower bounds for parameter estimation.

A method to improve LLMs by automating the construction of a mixture of expert prompts.

problem Limitation of single instruction prompts in covering complex problem spaces.
method Divide the problem space into sub-regions, each governed by a specialized expert with both an instruction and demos. A two-phase process constructs these experts.
result Achieves an average win rate of 81% across major benchmarks.

CurveRL optimizes large model reasoning by reweighting prompts based on their rank and density.

problem Improving large language model reasoning through context reweighting.
method Formulated prompt reweighting as a functional derivative, proposing CurveRL based on quantile coordinate transform.
result CurveRL consistently outperforms existing methods across multiple benchmarks.