AI conference papers

AI conference paper library

Search papers from 30+ top conferences, including NeurIPS, ICLR, ICML, CVPR, ACL and AAAI. Filter by keyword, then open any paper for its abstract, authors and links.

arXiv

Freeform Preference Learning for Robotic Manipulation

Marcel Torne, Anubha Mahajan, Abhijnya Bhat, Chelsea Finn

Reward design remains a central bottleneck for autonomous robot policy improvement, especially in long-horizon manipulation tasks where sparse success labels provide too little signal and binary preferences collapse many competing notions of quality into one ambiguous signal. We introduce Freeform Preference Learning (FPL), a method for learning robot policies from freeform human preferences. Rather than asking annotators which of two trajectories is better overall, FPL lets them define natural-language preference axes, such as speed, safety, quality of placement, or carefulness, and provide pairwise preferences along each axis. These annotations are used to learn a language-conditioned reward model that maps a trajectory and preference label to an axis-specific reward. We use this model to train a reward-conditioned policy that optimizes across the multiple human-specified dimensions. Across four real-world and two simulated long-horizon manipulation tasks, FPL improves over sparse-reward and binary-preference methods by 38 percentage points. Beyond improved performance, FPL learns dense progress signals without explicit subtask segmentation, shows compositionality of behavior not present in the data, and allows users to steer the policy towards different behaviors at test time without retraining. Blog post with videos available at https://freeform-pl.github.io/fpl.website/

arXiv

ApplE: A Modular Ontology of Applied Ethics and Event Context for Ethical Decision Modeling

Aisha Aijaz, Raghava Mutharaju, Manohar Kumar

Applied ethics applies ethical decision-making to domain-specific contexts using contextual information such as agents, actions, temporal and spatial settings, and theoretical constructs such as utility, virtues, rights, and duties. However, representing an ethical decision is challenging as it may be abstract, context-sensitive, and semantically heterogeneous. Nevertheless, important ethical and contextual factors can be formally modeled to support structured ethical reasoning. Knowledge representation and reasoning provide a mechanism to translate abstract ethical concepts into machine-interpretable conceptual structures in the context of an event. To achieve this, we propose ApplE, an Applied Ethics ontology that models ethical theory and event context within a unified and modular conceptual framework for ethical decision-making. The ontology was developed using a modified version of the Simplified Agile Methodology for Ontology Development (SAMOD), which facilitates iterative refinement of classes and relationships, as well as the participation of a domain expert. The modular development of ApplE combines Ethics Theory with Event Context to capture semantic relationships between ethical principles, agents, actions, consequences, intentions, and domains. Using ApplE, we modeled a use case from the medical domain to demonstrate the ontology's representational expressivity and reasoning capabilities. In addition to ontological reasoning and consistency checks, ApplE is also evaluated using the three-fold testing process of SAMOD. ApplE follows the FAIR principles and is positioned to be used as a reusable semantic and conceptual modeling resource for ethical AI systems and ontology-driven applications.

arXiv

Towards Anomaly Detection on Relational Data

Shiyuan Li, Yunfeng Zhao, Yue Tan, Qingfeng Chen, Yixin Liu, Shirui Pan

Relational databases are widely used for managing structured data in real-world systems. Detecting anomalies from such relational data is crucial for identifying fraud, risks, and abnormal behaviors, yet remains under-explored. The key challenges lie in the intrinsic complexity of relational data: multi-table attributes are high-dimensional and heterogeneous, making sparse abnormal clues easy to overwhelm by normal or irrelevant information. Moreover, anomalies may further manifest as abnormal connection patterns across different foreign-key relations, which existing tabular and graph anomaly detection methods are ill-suited to capture. To address them, we propose RelAD, a reconstruction-based framework that captures anomalies from both attribute and relational edge reconstruction. RelAD contains two core modules: conditional sparse-gated attribute reconstruction, which suppresses redundant multi-table attributes and emphasizes abnormal semantic blocks, and dual-view multi-relational edge reconstruction, which detects relation-specific abnormal connections from both intrinsic and behavioral instance profiles. The resulting attribute and relational signals are integrated through a lightweight fusion module to produce the final anomaly score. We further construct 6 benchmark datasets with systematic anomalies, on which extensive experiments show that RelAD consistently outperforms baselines while achieving competitive efficiency..The source code is available at https://github.com/Shiy-Li/RelAD.

arXiv

CARE: Context-Aware Ranking Evolution with Executable Scoring Programs for Budgeted Reaction Optimization

Guanyu Liu, Weiyi Kong, Chao Tang, Zeyu Wang, Boer Zhang, Baiqing Li, Peiyu Zhang, Tianyu Shi

High-throughput experimentation can evaluate many reaction conditions, yet combinatorial condition spaces still exceed the available experiment budget. This makes experiment selection a sequential decision problem: each new condition must be chosen from limited observations before its outcome is known. LLMs can express task-specific selection logic. A direct recommendation, however, is neither a persistent executable object that can be validated and revised nor an independently auditable decision rule. We introduce CARE, a reference-conditioned controller that separates program synthesis from experiment selection. An LLM writes an executable scoring program that ranks the remaining conditions, while a non-LLM reference policy supplies a numerical candidate and support summary. CARE forms an optional alternative from the program, applies a reference-conditioned intervention gate to compare it with the reference, and records the decision before the selected outcome is revealed. Each new outcome updates controller state and can trigger retention, revision, or regeneration of the active program. This outcome-guided program evolution changes the scoring logic without updating the LLM parameters. In matched offline replay with 30 seeds on eight reaction-optimization tasks, CARE attains the lowest normalized regret, the highest normalized best-so-far AUC, and the highest Top-1% Success@15 among the evaluated methods. These results support using a scoring program written by an LLM as one component of a reference-conditioned optimizer rather than as a standalone experiment selector.

arXiv

Using Seismic Statistical Features and VQ-VAE to Improve Spatiotemporal Seismicity Predictability

Wei Quan, Denise Gorse

In this paper we build upon a previous study in which we demonstrated, using XGBoost and earthquake catalogue data from Japan and Chile, that a set of 60 seismic statistical features (SSFs) had much greater predictive value than a set of 428 generic time series features from the tsfresh package. We here extend this previous work in two key ways, focusing on data from Japan as a large dataset is necessary in order to allow for the training of a deep learning (autoencoder) model. First, we move from whole-region prediction (considering, for each candidate event, the likelihood of an event M $\geq$ 5.0 anywhere in the region in the next 15 days) to localised predictions in which both the region of feature computation and the region of prediction are restricted to a circle of radius 24 km around the candidate event, and we show that performance remains excellent, similar to our previous whole-region study for the same area. Second, we here couple this proven set of SSFs, based on one-dimensional (catalogue) data, with a novel feature based on two-dimensional seismic maps, obtained by training a VQ-VAE model to reproduce such maps as output and identifying a measure of its error in doing so with a localised build-up of crustal stress. We show that while localised prediction based only on a set of SSFs can be effective, with test AUC values as high as those obtained in the case of Japan in our previous whole-region study, the inclusion of the new natively-spatial VQ-VAE-derived feature, top-ranked by SHAP analysis, provides valuable additional insights.

arXiv

TLA-Prover: Verifiable TLA+ Specification Synthesis via Preference-Optimized Low-Rank Adaptation

Eric Spencer, Arslan Bisharat, Brian Ortiz, Khushboo Bhadauria, Mujtaba Nazari, TaiNing Wang, George K Thiruvathukal, Konstantin Laufer, Mohammed Abuhamad

TLA+ is a formal specification language for verifying distributed systems and safety-critical protocols. Large language models (LLMs) frequently produce TLA+ specifications that fail the TLC model checker for semantic reasons. Across 25 LLMs, the best public baseline is 26.6% syntactic parse and 8.6% semantic model-check. We present TLA-Prover, a 20-billion-parameter model for TLA+ specification synthesis. Training combines supervised fine-tuning (SFT) on verified examples with repair-based group-relative policy optimization (GRPO). In the GRPO stage, the model learns to fix its own rejected specifications. We also train a direct preference optimization (DPO) variant from the same SFT checkpoint as an ablation. TLC provides the reward signal directly, with no learned reward model. Four tiers grade each output: Bronze (parses), Silver (no warnings), Gold (passes TLC), and Diamond. To reach Diamond, the model's correctness property is automatically altered in a small way; TLC must then detect a violation. If TLC still passes, the property was always-true and contributes nothing; the output fails Diamond. TLA-Prover reaches 9/30 (i.e. pass@1 = 30%) at both Gold and Diamond on a held-out 30-problem benchmark. This is roughly 3.5x the 8.6% untuned baseline. The DPO variant reaches 20% at Diamond. Gold and Diamond coincide at every checkpoint; this prevents the trivial-property failure mode.

arXiv

GeoFlowVLM: Geometry-Aware Joint Uncertainty for Frozen Vision-Language Embedding

Mayank Nautiyal, Li Ju, Andreas Hellander, Ekta Vats, Prashant Singh

Standard dual-encoder vision-language models that map images and text to deterministic points on a shared unit hypersphere through $\ell_2$ normalization typically expose neither \emph{aleatoric} uncertainty (cross-modal ambiguity) nor \emph{epistemic} uncertainty (lack of training-distribution support). Existing post-hoc methods either recover at most one of the two uncertainty components, or ignore the hyperspherical geometry of these models' embeddings. We propose \textbf{GeoFlowVLM} as a post-hoc adapter that learns the joint distribution of paired $\ell_2$-normalised dual-encoder VLM embeddings on the product hypersphere $\mathbb{S}^{d-1} \times \mathbb{S}^{d-1}$ via Riemannian flow matching with a single masked velocity field. A consistency result shows that, in the population limit, the trained network exposes the joint flow and both cross-modal conditional flows as valid Riemannian flow-matching velocity fields on their respective domains. We derive two quantities from this single model: a conditional retrieval entropy that quantifies aleatoric ambiguity with a decision-theoretic interpretation via a Fano-type bound, and a marginal-typicality epistemic score justified by an exact chain-rule decomposition of the joint NLL. This decomposition isolates a cross-modal pointwise-mutual-information term that is structurally discriminative rather than epistemic, and is empirically the only consistently uninformative standalone component. Empirically, the entropy tracks Recall@1 with near-ideal monotonic calibration across three retrieval benchmarks in both directions, and the marginal-typicality sum yields consistently calibrated selective accuracy across four zero-shot classification benchmarks.

arXiv

Optimal Regret for Single Index Bandits

Devdan Dey, Sujoy Bhore, Avishek Ghosh

We study the $\textit{single-index bandit}$ problem, where rewards depend on an unknown one-dimensional projection of high-dimensional contexts through an unknown reward function. This model extends linear and generalized linear bandits to a nonparametric setting, and is particularly relevant when the reward function is not known in advance. While optimal regret guarantees are known for monotone reward functions, the general non-monotone case remains poorly understood, with the best known bound being $\tilde{\mathcal{O}}(T^{3/4})$ (under standard boundedness and Lipschitz assumptions on the reward function [Kang et al., 2025]). We close this gap by establishing the optimal regret for general single-index bandits. We propose a simple two-phase algorithm, namely, Zoomed Single Index Bandit with Upper Confidence Bound ($\texttt{ZoomSIB-UCB}$), that first estimates the projection direction via a normalized Stein estimator, and then reduces the problem to a one-dimensional bandit using discretization and finally use UCB. This approach achieves a regret of $\tilde{\mathcal{O}}(T^{2/3})$, and improves significantly upon prior work without any additional assumptions. We also prove a matching minimax lower bound of $\tildeΩ(T^{2/3})$, showing that the upper bound is essentially tight. Our upper and lower bounds together provide a sharp characterization of the regret in single-index bandits. Moreover, the empirical results further demonstrate the effectiveness and robustness of our approach.

arXiv

Tail-Aware Information-Theoretic Bounds for LLM Alignment under Heavy-Tailed Rewards

Huiming Zhang, Binghan Li, Wan Tian, Qiang Sun

Classical information-theoretic learning bounds typically rely on KL mutual information and moment-generating-function (MGF) arguments, which are well matched to bounded or sub-Gaussian losses but can be ineffective when losses or rewards are heavy-tailed. We develop a tail-aware information-theoretic framework for sub-Weibull data, where the tail parameter $θ$ controls the tail heaviness: $θ=2$ corresponds to sub-Gaussian, $θ=1$ to sub-exponential, and $0<θ<1$ to genuinely heavy tails. Our key technical ingredient is a decorrelation lemma that bounds change-of-measure expectations using a shifted-log $f_θ$-divergence, which admits explicit comparisons to Rényi divergence without MGF arguments. On the empirical-process side, we establish sharp maximal inequalities and a Dudley-type chaining bound for sub-Weibull processes, with logarithmic and entropy terms raised to the power $1/θ$. These tools yield tail-adaptive selection bounds and a multiscale information-theoretic Dudley inequality based on shifted-log and Rényi mutual information. We apply our theory to large language models (LLMs) in the context of reward hacking within reinforcement learning from human feedback (RLHF). We show that Rényi-regularized alignment provides finite reward guarantees and ensures that best-of-N policies remain well-controlled, thereby mitigating the catastrophic Goodhart effects where standard KL-regularization fails. We illustrate Rényi-regularized RLHF by experiments, including controlled heavy-tailed rewards and token-space reward attacks.

arXiv

Stochastic-Dimension Frozen Sampled Neural Network for High-Dimensional Gross-Pitaevskii Equations on Unbounded Domains

Zhangyong Liang

This paper introduces the Stochastic-Dimension Frozen Sampled Neural Network (SD-FSNN), a novel computational framework for solving high-dimensional Gross-Pitaevskii equation (GPE) on unbounded domain. The proposed method circumvents the curse-of-dimensionality that plagues traditional discretizations and the computational bottlenecks of gradient-based neural network solvers through a synergistic combination of techniques. First, a prescribed Gaussian envelope encodes the far-field decay of the wavefunction, enabling a space-time separation where the spatial approximation is handled by a frozen, single-hidden-layer neural network with data-driven sampled features. This yields a gradient-free formalism where spatial derivatives are analytically precomputed and time-dependence is evolved via reduced ODEs. Second, a stochastic-dimension sampler provides a conditionally unbiased estimate of the spatial operator by evaluating only a small subset of spatial dimensions at each time step, essentially reducing computational and memory costs. Discrete conservation laws are also enforced, ensuring long-term stability. Extensive numerical experiments on GPE in up to 1000 dimensions demonstrate that SD-FSNN achieves significantly higher accuracy and efficiency compared to state-of-the-art methods, including PINNs, randomized feature methods, and tensor-network approaches. The results confirm that SD-FSNN effectively mitigates the Kolmogorov $n$-width barrier for frozen-basis models on structured solution manifolds.

arXiv

SIGMA: Semantic Identifier Grouping for Molecular Autoregression

Xinyu Wang, Fei Dou, Jinbo Bi, Minghu Song

Autoregressive molecular models assign probability to molecular serializations even though chemical identity is invariant to serialization. Equivalent serializations can therefore represent a common molecular identity yet induce inconsistent next-token decisions. Randomized strings broaden exposure, but do not reveal which intermediate decisions should agree. We introduce SIGMA, a dense suffix-position objective built from chemically certified same-suffix triplets: two equivalent histories, one non-equivalent history, and a shared suffix. SIGMA aligns corresponding pre-token hidden states along the continuation and separates the negative to a finite relative margin, leaving the language-model objective, decoder, and inference procedure unchanged. We compare SIGMA with canonical training, randomized-serialization training, and last-token alignment across four datasets under SMILES and SELFIES. Across the eight representation-dataset blocks, SIGMA yields clear test-reference Frechet ChemNet Distance reductions in six: all four SELFIES domains and QM9 and ZINC under SMILES, with paired 95% confidence intervals below zero against every control. Position-wise analyses show improved state correspondence, chemical discrimination, and next-token agreement while preserving between-molecule information. On two full-corpus ZINC blocks, compute-matched ablations identify chemically correct state correspondence as the effective ingredient. Beyond generation, SIGMA improves mean predictive performance on all six molecular property benchmarks and reduces sensitivity to equivalent molecular serializations on every task.

arXiv

When Differential Privacy Meets Wireless Federated Learning: An Improved Analysis for Privacy and Convergence

Chen Yaoling, Liang Hao, Tu Xiaotong

Differentially private wireless federated learning (DPWFL) is a promising framework for protecting sensitive user data. However, foundational questions on how to precisely characterize privacy loss remain open, and existing work is further limited by convergence analyses that rely on restrictive convexity assumptions or ignore the effect of gradient clipping. To overcome these issues, we present a comprehensive analysis of privacy and convergence for DPWFL with general smooth non-convex loss objectives. Our analysis explicitly incorporates both device selection and mini-batch sampling, and shows that the privacy loss can converge to a constant rather than diverge with the number of iterations. Moreover, we establish convergence guarantees with gradient clipping and derive an explicit privacy-utility trade-off. Numerical results validate our theoretical findings.

arXiv

Local Shapley: Model-Induced Locality and Optimal Reuse in Data Valuation

Xuan Yang, Hsi-Wen Chen, Ming-Syan Chen, Jian Pei

The Shapley value provides a principled foundation for data valuation, but exact computation is #P-hard due to the exponential coalition space. Existing accelerations remain global and ignore a structural property of modern predictors: for a given test instance, only a small subset of training points influences the prediction. We formalize this model-induced locality through support sets defined by the model's computational pathway (e.g., neighbors in KNN, leaves in trees, receptive fields in GNNs), showing that Shapley computation can be projected onto these supports without loss when locality is exact. This reframes Shapley evaluation as a structured data processing problem over overlapping support-induced subset families rather than exhaustive coalition enumeration. We prove that the intrinsic complexity of Local Shapley is governed by the number of distinct influential subsets, establishing an information-theoretic lower bound on retraining operations. Guided by this result, we propose LSMR (Local Shapley via Model Reuse), an optimal subset-centric algorithm that trains each influential subset exactly once via support mapping and pivot scheduling. For larger supports, we develop LSMR-A, a reuse-aware Monte Carlo estimator that remains unbiased with exponential concentration, with runtime determined by the number of distinct sampled subsets rather than total draws. Experiments across multiple model families demonstrate substantial retraining reductions and speedups while preserving high valuation fidelity.

arXiv

Impact of Dataset Composition on Embedded Real-Time UAV Wildfire Detection Using Compact YOLO Models

Eduardo de los Santos, Andre S. Kelbouscas, Ricardo B. Grando, Bruna V. Guterres

The development of vision-based wildfire detection systems for unmanned aerial vehicles is constrained by the limited availability of diverse real-world training images. This paper investigates the impact of dataset composition on embedded real-time UAV wildfire detection using compact YOLO models as a controlled validation family. Four training configurations were evaluated: real non-augmented, real augmented, hybrid non-augmented, and hybrid augmented, where the hybrid sets combine real wildfire images with AI-generated samples. The objective is to determine whether synthetic data mixing and image augmentation improve practical detection performance under resource-constrained deployment conditions. Experimental results show that the best overall operating point was obtained with the real non-augmented dataset, which achieved the strongest balance between recall and mean average precision for UAV-based wildfire detection. The results also show that neither hybridization with synthetic data nor augmentation produced a better final deployment choice. These findings suggest that, for embedded UAV wildfire detection, dataset realism and domain alignment are more valuable than increasing training set size through synthetic expansion.

arXiv

Auditing Medical Vision-Language Models on Chest Radiographs: Estimating Reference Agreement Across Institutions

Pengyang Yu, Yiou Wang, Zhongping Dong, Sahraoui Dhelim, Chun-Mei Feng, M. Tahar Kechadi

Vision-language models return structured chest-radiograph findings through interfaces exposing no confidence score, so a receiving institution cannot read off how far to trust an individual judgment. Whether agreement with an institution's reference standard transfers across sites, findings, prediction directions and question formats is largely unmeasured. We evaluated three generative vision-language models on three institutional chest-radiograph corpora and six findings under two elicitation protocols, comprising more than 345,000 finding-level predictions, and estimated finding-by-direction reference agreement at a receiving institution from a small budget of local labels. Estimation strategies were then stress-tested under repeated strict institution-held-out evaluation. Under evaluation excluding the receiving institution from development entirely, adaptive selection among the seven estimators that design admits did not improve on simple fixed alternatives: it achieved a mean Brier score of 0.1083, against 0.0853 for always using a Beta-Binomial empirical-Bayes estimator and 0.0855 for a target-only logistic model. Those two differ by 0.0003, less than this family's own sensitivity to a change of solver version, and each leads in about half the settings, so no default can be recommended. Their advantage over estimators pooling across institutions was concentrated at one site and not confirmatory once clustered by institution, and a plug-in empirical-Bayes posterior-predictive count interval at a nominal 95% level covered 87.0%, less at the hardest institution. Reference agreement therefore has to be re-evaluated per site and per interface; these results concern agreement with institutional labels, not clinical correctness.

arXiv

P2Voxel: Pyramid Pivot Voxelization for 3D Mesh Tokenization

Zhenhong Sun, Haozhe Liu, Yifu Wang, Xibin Song, Senbo Wang, Huadong Mo, Daoyi Dong, Hongdong Li, Pan Ji

Triangle meshes provide explicit and accurate surface geometry, yet their irregular topology connectivity makes 3D mesh tokenization a geometric sampling problem: how to sample and organize geometric evidence into compact, structured and learnable tokens. Beyond field-centric volumetric sampling and edge-intersection surface sampling, we retarget mesh tokenization as \textit{local surface evidence sampling}: identifying the minimal geometric evidence inside each active voxel that is sufficient for deterministic surface recovery. To this end, we introduce \textbf{P2Voxel}, a pyramid pivot voxelization framework for compact and reconstruction-aware mesh tokenization. P2Voxel is built on three key innovations. Under the \textit{Local Planarity} assumption, Pivot Voxelization represents each active voxel with a surface pivot and an orientation sign, providing minimal local evidence that can induce the corner values required for deterministic reconstruction. Under the \textit{Spatial Complexity} assumption, Pyramid Pivot Voxelization exploits the spatial non-uniformity of real surfaces by allocating finer pivot tokens to geometrically complex regions while keeping smooth regions coarse and compact. Under the \textit{Block Reconstructability} assumption, a Pyramid VAE learns compact multi-resolution latent codes over locally reconstructable pivot blocks, avoiding the need to model the entire high-resolution voxelized shape as a dense global field. Together, these designs convert meshes into compact, structured, and learnable pyramid pivot tokens, enabling efficient mesh reconstruction for downstream 3D tasks.

arXiv

Understanding Rate-Distortion Performance in Distributed Transformer Inference

Anderson de Andrade, Alon Harell, Ivan V. Bajić

Transformers achieve superior performance on many tasks, but impose heavy compute and memory requirements during inference. This inference can be made more efficient by partitioning the process across multiple devices, which, in turn, requires compressing its intermediate representations. We study compressibility of transformer's intermediate representations via learned compression through the lens of rate-distortion and the theory of usable information. Our study reveals that, unlike convolutional models, deeper representations in transformers become more difficult to compress. The reason for this behavior is twofold: first, the complexity of representation increases as we move deeper into the transformer, necessitating a higher rate; and second, the higher complexity of representations worsens the generaliztion bound for learned entropy estimates, which further compromises compression performance. Through a combination of experiments and theory, we characterize and analyze the compressibility of transformer representations, derive bounds on the achievable rate of learned codecs applied to these representations, and offer a unified lens for understanding rate-distortion performance in representation coding.

arXiv

Rethinking Federated Graph Foundation Models: A Graph-Language Alignment-based Approach

Yinlin Zhu, Di Wu, Xianzhi Zhang, Yuming Ai, Xunkai Li, Miao Hu, Guocong Quan

Recent studies of federated graph foundational models (FedGFMs) break the idealized and untenable assumption of having centralized data storage to train graph foundation models, and accommodate the reality of distributed, privacy-restricted data silos. Despite their simplicity and intuition, existing studies that project aligned generalizable knowledge onto a discrete token space via vector-quantized backbones suffer from irreversible knowledge loss during the quantization process. In this context, we argue that reconciling the semantic-structural orthogonality and integrity between pre-trained language models (PLMs) and graph neural networks (GNNs) is paramount for developing effective FedGFMs while simultaneously mitigating the severe data heterogeneity and communication constraints inherent in distributed, resource-limited environments. To address these issues, we propose FedGALA (Federated Graph And Language Alignment), a framework that resolves graph-based semantic-structural orthogonality and integrity in federated settings by employing unsupervised contrastive learning to align GNNs and frozen PLMs within a continuous embedding space, thereby capturing robust, transferable general knowledge. Subsequently, FedGALA leverages a communication-efficient prompt tuning mechanism to steer these pre-aligned encoders and frozen PLMs, facilitating effective adaptation to diverse downstream tasks while circumventing the prohibitive overhead of full-parameter fine-tuning. The comprehensive experiments validate that FedGALA outperforms all competitive baselines across multi-domain datasets on multiple tasks with up to 14.37% performance improvement.

arXiv

Turn-Based Structural Triggers: Structure-Conditioned Backdoors in Multi-Turn LLMs

Yiyang Lu, Jinwen He, Yue Zhao, Kai Chen, Ruigang Liang, Cheng Hong, Yingjun Zhang

Large Language Models (LLMs) are increasingly deployed as multi-turn assistants and customized through instruction tuning with project-specific training components. This practice creates a supply-chain risk when organizations reuse third-party fine-tuning frameworks, trainer extensions, or outsourced training code: an adversary who subtly compromises the loss-computation component can inject malicious supervision during fine-tuning while leaving the stored training corpus, model architecture, and deployment interface unchanged. Existing LLM backdoors and defenses are largely prompt-centric, relying on lexical, syntactic, or semantic patterns in user inputs while overlooking structural signals in multi-turn conversations. We propose Turn-based Structural Trigger (TST), a prompt-free backdoor that uses dialogue turn position as its activation condition. TST exploits structural cues implicitly encoded by chat templates and is implanted without modifying the stored dialogue corpus. Its trigger is automatically present once the conversation reaches the attacker-specified turn, making activation independent of downstream user inputs and resistant to prompt filtering, sanitization, and paraphrasing. In our primary setting, the model behaves normally during early interactions and activates the attacker-defined behavior only after the conversation reaches the designated structural condition. Across four open-source LLM families, TST achieves an average Attack Success Rate of 98.10% on target turns and a Clean Rate of 99.96% on non-target turns, while retaining 97.78% of clean-model utility. These results identify dialogue structure as an overlooked attack surface and motivate structure-aware auditing beyond prompt inspection.

arXiv

Robust Bayesian Optimization via Tempered Posteriors

Jiguang Li, Hengrui Luo

Bayesian optimization (BO) iteratively fits a Gaussian process (GP) surrogate to accumulated evaluations and selects new queries via an acquisition function. Under local misspecification, this feedback loop can produce overconfidence precisely in the region guiding subsequent decisions. We develop a tempered GP-based BO framework that raises the likelihood to a power $α\in(0,1]$. For a generalized family of improvement acquisitions indexed by $g$, including probability of improvement (PI, $g=0$) and expected improvement (EI, $g=1$), we derive finite-time cumulative regret bounds with adaptively learned kernel hyperparameters. The analysis shows that tempering reduces the noise-driven confidence and information-gain contributions to regret, while a deterministic RKHS term prevents arbitrarily aggressive tempering from being uniformly beneficial. It also clarifies the role of the acquisition function: positive-order $g$-EI rules preserve the usual information-gain regret behavior, whereas zero-jitter PI is more exploitative and admits a weaker worst-case guarantee. Motivated by our theoretic findings, we propose a prequential procedure for selecting $α$ online: it decreases $α$ when realized prediction errors exceed model-implied uncertainty and returns $α$ toward one as calibration improves. Empirical results demonstrate that tempering provides a practical yet theoretically grounded tool for stabilizing BO surrogates under localized sampling.

Keep browsing

More papers from top AI conferences

30+ top conferences, by conference, year, and topic.