arXiv subjects

Sunwoo Kim

Publications and source records attributed to Sunwoo Kim.

At least 19 recordsLinked to original sources

arXiv2026

The Role of Fine-grained Harm Signals in LLM Safety

Prior work has shown that internal harmfulness representations in large language models vary across risk categories, while sharing a common general harm representation component. This raises a question about the role of the category-specific component beyond general harm representation in LLM safety. To answer this question, we isolate the category-specific component by removing shared general harmfulness representation from each categorical harmfulness representation, yielding a category residual that is orthogonal to general harmfulness at every layer. Using activation steering with category residuals across 11 risk categories in 3 instruction-tuned LLMs, we find that whether category residuals encode harmfulness varies across categories, and that this category-wise pattern is similar across models. Whether category residuals induce refusal also varies across categories, but this category-wise pattern is more model-dependent. We also find that category residuals increase LLMs' downstream internal alignment with shared general harmfulness representation. Together, these findings demonstrate that more fine-grained category residuals should also be considered beyond shared general harmfulness representation to fully understand LLM safety. More broadly, our findings show that even a direction orthogonal to a concept at one layer can contribute to the concept's downstream amplification.

cs.CL↗

arXiv2026

Form Over Content In Gradient-Based Data Attribution Methods

Data attribution methods using gradient similarity are widely used to analyze and select training data for large language models, but what gradient similarity actually measures is debated. Some interpret it as identifying task-relevant skills, while other work reports that surface form is the main factor. We resolve this debate for supervised fine-tuning examples by varying task and answer format independently. Specifically, we render benchmarks in different answer formats, such that datasets can share a task without a format or a format without a task. We find that gradient alignment follows the answer format, as benchmark pairs sharing an answer format align strongly (disattenuated cosine near 0.4), while same benchmarks rendered with different answer format classes show no alignment (near 0.0). We demonstrate that this ordering holds from the earliest pretraining checkpoints through post-training, and across model scales and families. We then analyze the released selections of LESS, a gradient-based data selection method for instruction tuning, and find that each target's selections over-represent the target's own answer format. Hence, we demonstrate that gradient-based attribution methods track format similarity more than task semantics, meaning that such methods, as well as the semantic interpretation of the gradient, should be tested on data where answer format and task vary independently for greater robustness and reliability.

cs.CL↗

arXiv2026

The Resurrection of Spectrum Spreading for 6G and Beyond: From Sinusoids to Chirps

Orthogonal frequency-division multiplexing (OFDM) and its sinusoidal subcarriers have underpinned the 4G and 5G eras, delivering high spectral efficiency and resilience to multipath fading through an efficient multicarrier architecture. However, as future systems move toward doubly dispersive environments driven by high-mobility applications and migration to mmWave/sub-THz bands, the time-invariance assumption underlying OFDM becomes increasingly difficult to maintain, and Doppler-induced degradation becomes prominent. While enhancements such as MIMO, advanced coding, and scheduling provide incremental remedies, they introduce additional overhead, because the sinusoidal subcarrier itself offers no inherent waveform-level robustness to Doppler impairments. Accordingly, two time-frequency spreading philosophies have emerged to improve Doppler resilience by distributing each symbol's energy across both dimensions of the time-frequency plane: (i) 2D isotropic spreading via the delay-Doppler (DD) domain, exemplified by the orthogonal time frequency space (OTFS) family, and (ii) sheared spreading via parameterizable chirps, exemplified by the affine frequency-division multiplexing (AFDM) family. In this article, we examine key considerations for future waveform design across these paradigms and argue that transitioning from the sinusoidal subcarriers of OFDM to the chirp-based subcarriers offers a viable design direction for improving Doppler robustness while retaining much of the mature OFDM infrastructure. This perspective also highlights the suitability of chirp-based waveforms for integrated sensing and communications (ISAC) and their extensibility to emerging physical-layer techniques. Overall, we argue that the transition from sinusoids to chirps is a technically motivated, compelling evolutionary direction for future wireless physical layer design.

eess.SP↗

arXiv2026

BeamGuard: Risk-Aware Multimodal Beam Forecasting and Adaptive Virtual Beamwidth Control for 6G mmWave V2I Links

Reliable beam management is a central challenge for 6G millimeter-wave (mmWave) vehicle-to-infrastructure (V2I) links, where narrow beams provide high array gain but are vulnerable to mobility-induced misalignment, blockage, and domain variation. BeamGuard is a multimodal sensing-aided beam-management framework that combines exteroceptive sensing with optional partial in-band mmWave power observations to forecast future beam distributions and select adaptive virtual beamwidth actions for reliable V2I control. It fuses camera, radar, LiDAR, GPS, and mmWave power observations with a temporal multimodal forecaster, then converts the predicted posterior into a beam center and virtual codebook-level beamwidth through a risk-aware planner. Here, virtual beamwidth denotes adjacent-beam coverage in the codebook index space rather than physical analog wide-beam synthesis. BeamGuard supports sensor-only operation for beam-training overhead reduction, limited in-band operation with masked beam-power entries, and full hybrid operation with sensing and communication-side measurements. We evaluate BeamGuard on DeepSense 6G Scenarios 32 and 33, with additional held-out tests on Scenarios 31 and 34, covering day--night training, transfer, limited adaptation, ablations, budget sweeps, and lightweight baselines. The full-hybrid anchor, used as the complete-system reference, achieves Top-1/Top-3/Top-5 accuracies of approximately $0.393/0.778/0.897$, while the planner attains a threshold-based outage probability of about $0.0060$ with a gain ratio of about $0.895$. Matched-budget baselines further show that BeamGuard improves over multilayer perceptron, recurrent, and temporal convolutional predictors under comparable in-band observation settings. These results demonstrate robust, overhead-aware beam management through multimodal forecasting and risk-aware virtual beamwidth control.

eess.SP↗

arXiv2026

Wontopos Tablet 2: Measuring Multilingual and Multimodal Memory Retrieval Without Lexical Matching

We measure tablet-2, a production long-term memory engine for language models, on the text benchmarks the field already uses and on cross-lingual retrieval of photographs stored with no text at all. Its retrieval path contains no lexical matching, no keyword scoring, and no language model of its own. On LongMemEval-S (500 questions) it scores 95.7% [93.4, 97.1]; on BEAM-1M (700 questions, 2.21M stored memories) 67.5% [64.8, 70.2]. Those are question-sampling intervals, not the run-to-run spread, which is an order of magnitude narrower. Most of the paper is about how little they mean alone. Holding engine, corpus, settings and judge fixed, changing only the reader moves LongMemEval-S by 2.0 points; changing only the re-ask budget moves BEAM-1M by 8.9. Neither is stated in the reports we compare against, and the second exceeds most gaps there, so we give that table as a placement and not a ranking. For the multimodal axis we run two controls. Against BM25, configured as strongly as we could, we reach 95.2% mean recall@5 over 70 store-and-query language cells where BM25 reaches 19.0% and is exactly zero in 54. On captionless photographs a lexical method has no document to score at all. Open dense baselines on 300 Crossmodal-3600 photographs in 14 languages show that density confers no language independence: one scores 91.0% on English and 4.7% on Russian from identical image vectors, and a multilingual variant collapses on Telugu and Swahili. Our spread across languages is 14.0 against their 27.5 and 27.7. Three results run against us and are reported at equal weight: low-resource languages degrade sharply (Swahili 53.0%, Telugu 64.0%), attaching captions lowers cross-lingual retrieval by 11.4 points, and one setting omitted into one stage of our own retrieval cost 37 points of Korean top-1 accuracy while leaving nine languages untouched.

cs.IR↗

arXiv2026

Training-Free LLM-Based Recommendation with Post-LLM Item Refinement Using Collaborative Signals

Large language models (LLMs) have shown promise for training-free recommendation, but LLM-generated user interests are often too broad for fine-grained item retrieval. Existing methods incorporate collaborative filtering (CF) signals in a pre-LLM manner through candidate reranking or prompt augmentation, yielding limited gains. We propose CoRRe, a training-free recommendation framework with a post-LLM paradigm that injects CF signals into LLM-generated item representations, which are later matched with LLM-generated user interests for ranking. Specifically, CoRRe refines the directions of item embeddings using an item-item co-purchase graph and their magnitudes using item popularity. Experiments on real-world datasets show that CoRRe consistently outperforms existing training-free methods and achieves competitive or superior performance compared with training-based methods, without requiring any model training or task-specific fine-tuning.

cs.IR↗

arXiv2026

Multi-Chirp AFDM for Rydberg Atomic Quantum Receivers: Waveform and Algorithm Design

We propose a multi-chirp affine frequency division multiplexing (MC-AFDM) scheme for joint delay-Doppler estimation with Rydberg atomic quantum receivers (RAQRs). The work is motivated by the fact that RAQRs, while offering superior sensitivity and advantageous sensing capabilities, suffer from an optical ambiguity due to Doppler shifts in doubly-dispersive (DD) channel caused by target mobility, which precludes the reliable estimation of delay-Doppler parameters. To resolve this optical ambiguity and unleash the potential of RAQRs in DD channel, the proposed MC-AFDM employs multiple distinct AFDM post-chirp signals to overcome the rank-deficiency problem of the classical single-chirp AFDM (SC-AFDM), thereby enabling accurate delay-Doppler estimation of multiple targets. Our analysis reveals that the edge distribution of the multiple post-chirp parameters can further improve estimation accuracy by minimizing the condition number. Building on the proposed MC-AFDM waveform, we design a sequential signal processing algorithm based on orthogonal matching pursuit (OMP) and least squares (LS), and we derive the theoretical lower bounds for delay and Doppler estimation. Numerical results show that the proposed MC-AFDM improves range and velocity estimation accuracy by up to two orders of magnitude compared to the classical SC-AFDM, and approaches its theoretical bounds through post-chirp optimization, validating the quantum-induced advantage of RAQRs for high-resolution quantum wireless sensing.

eess.SP↗

arXiv2026

PapersPlease: A Benchmark for Evaluating Motivational Values of Large Language Models Based on ERG Theory

Evaluating the performance and biases of large language models (LLMs) through role-playing scenarios is becoming increasingly common, as LLMs often exhibit biased behaviors in these contexts. Building on this line of research, we introduce PapersPlease, a benchmark consisting of 3,700 moral dilemmas designed to investigate LLMs' decision-making in prioritizing various levels of human needs. In our setup, LLMs act as immigration inspectors deciding whether to approve or deny entry based on the short narratives of people. These narratives are constructed using the Existence, Relatedness, and Growth (ERG) theory, which categorizes human needs into three hierarchical levels. Our analysis of six LLMs reveals statistically significant patterns in decision-making, suggesting that LLMs encode implicit preferences. Additionally, our evaluation of the impact of incorporating social identities into the narratives shows varying responsiveness based on both motivational needs and identity cues, with some models exhibiting higher denial rates for marginalized identities. All data is publicly available at https://github.com/yeonsuuuu28/papers-please.

cs.CL↗

arXiv2026

On the Effect of Uncertainty on Layer-wise Inference Dynamics

Understanding how large language models (LLMs) internally represent and process their predictions is central to detecting uncertainty and preventing hallucinations. While several studies have shown that models encode uncertainty in their hidden states, it is underexplored how this affects the way they process such hidden states. In this work, we demonstrate that the dynamics of output token probabilities across layers for certain and uncertain outputs are largely aligned, revealing that uncertainty does not seem to affect inference dynamics. Specifically, we use the Tuned Lens, a variant of the Logit Lens, to analyze the layer-wise probability trajectories of final prediction tokens across 11 datasets and 5 models. Using incorrect predictions as those with higher epistemic uncertainty, our results show aligned trajectories for certain and uncertain predictions that both observe abrupt increases in confidence at similar layers. We balance this finding by showing evidence that more competent models may learn to process uncertainty differently. Our findings challenge the feasibility of leveraging simplistic methods for detecting uncertainty at inference. More broadly, our work demonstrates how interpretability methods may be used to investigate the way uncertainty affects inference.

cs.CL↗

arXiv2026

On the Memorization Behavior of LLMs in Generative Recommendation: Observations, Implications, and Training Strategies

Generative recommendation (GR) has emerged as a promising direction for recommender systems. Recently, large language models (LLMs) have been increasingly adopted for GR, as their rich pretrained knowledge is expected to help them generalize beyond common user behavior patterns that traditional memorization-oriented baselines can capture. However, existing LLM-based GR works largely ignore LLMs' well-known tendency to memorize, which, if present in LLMs fine-tuned for GR, would restrict their utilization of pretrained knowledge. In this work, we investigate this concern by examining one-hop memorization, where a model recommends items that are direct successors of items in the training data. We show that LLMs do this more than non-LLM-based GR models-in fact, the vast majority of their gains over GR baselines are actually on users whose target items can be predicted through one-hop memorization. We intuit that improving performance on the remaining users requires LLMs to learn richer item-item relations beyond one-hop transitions. To achieve this, we propose IIRG, a novel training strategy that teaches LLMs to capture: (1) collaborative relations derived from item co-occurrences across multiple hops in user sequences, and (2) semantic relations among items with similar themes, both of which can serve as useful recommendation signals. We show that IIRG significantly improves over LLMs trained solely with standard next-item prediction, with especially large gains for users whose test items are not covered by train-time one-hop transitions.

cs.IR↗

arXiv2026

Rethinking Contrastive Learning for Graph Collaborative Filtering: Limitations and a Simple Remedy

Graph collaborative filtering (GCF) is a dominant paradigm in recommender systems, where contrastive learning (CL) objectives such as the Sampled Softmax (SSM) loss are widely used for optimization. However, it remains unclear how CL interacts with the prediction mechanism of GCF. By unfolding the prediction mechanism of GCF, we show that the user-item prediction score is computed by aggregating learnable weights over a large number of neighbor pairs formed by the multi-hop neighbors of the user and the item. This analysis suggests that effective optimization critically depends on which neighbor pairs are upweighted during training. Empirically, we find that effective recommendation is achievable by selectively upweighting only a small subset of neighbor pairs whose constituent neighbors are structurally similar to the target user and item, and that the effect of such selective upweighting varies across different neighbor pair types. Based on these findings, we analyze SSM and identify key limitations in its neighbor pair weight update dynamics. To address these limitations, we propose NT-SSM, an effective and principled CL objective that induces type-aware neighbor pair weight update dynamics. Experiments demonstrate consistent performance improvements over SSM across multiple datasets and GCF models.

cs.IR↗

arXiv2026

ItemRAG: Item-Based Retrieval-Augmented Generation for LLM-Based Recommendation

Recently, large language models (LLMs) have been widely used as recommender systems, owing to their reasoning capability and effectiveness in handling cold-start items. A common approach prompts an LLM with a target user's purchase history to recommend items from a candidate set, often enhanced with retrieval-augmented generation (RAG). Most existing RAG approaches retrieve purchase histories of users similar to the target user; however, these histories often contain noisy or weakly relevant information and provide little or no useful information for candidate items. To address these limitations, we propose ItemRAG, a novel RAG approach that shifts focus from coarse user-history retrieval to fine-grained item-level retrieval. ItemRAG augments the description of each item in the target user's history or the candidate set by retrieving items relevant to each. To retrieve items not merely semantically similar but informative for recommendation, ItemRAG leverages co-purchase information alongside semantic information. Especially, through their careful combination, ItemRAG prioritizes more informative retrievals and also benefits cold-start items. Through extensive experiments, we demonstrate that ItemRAG consistently outperforms existing RAG approaches under both standard and cold-start item recommendation settings. Supplementary materials, code, and datasets are provided at https://github.com/kswoo97/ItemRAG.

cs.IR↗

arXiv2026

Occlusion-Aware Multimodal Beam Prediction and Pose Estimation for mmWave V2I

We propose an occlusion-aware multimodal learning framework that is inspired by simultaneous localization and mapping (SLAM) concepts for trajectory interpretation and pose prediction. Targeting mmWave vehicle-to-infrastructure (V2I) beam management under dynamic blockage, our Transformer-based fusion network ingests synchronized RGB images, LiDAR point clouds, radar range-angle maps, GNSS, and short-term mmWave power history. It jointly predicts the receive beam index, blockage probability, and 2D position using labels automatically derived from 64-beam sweep power vectors, while an offline LiDAR map enables SLAM-style trajectory visualization. On the 60 GHz DeepSense 6G Scenario 31 dataset, the model achieves 50.92\% Top-1 and 86.50\% Top-3 beam accuracy with 0.018 bits/s/Hz spectral-efficiency loss, 63.35\% blocked-class F1, and 1.33m position RMSE. Multimodal fusion outperforms radio-only and strong camera-only baselines, showing the value of coupling perception and communication for future 6G V2I systems.

eess.SP↗

arXiv2026

Sensing-Assisted Adaptive Beam Probing with Calibrated Multimodal Priors and Uncertainty-Aware Scheduling

Highly directional mmWave/THz links require rapid beam alignment, yet exhaustive codebook sweeps incur prohibitive training overhead. This letter proposes a sensing-assisted adaptive probing policy that maps multimodal sensing (radar/LiDAR/camera) to a calibrated prior over beams, predicts per-beam reward with a deep Q-ensemble whose disagreement serves as a practical epistemic-uncertainty proxy, and schedules a small probe set using a Prior-Q upper-confidence score. The probing budget is adapted from prior entropy, explicitly coupling sensing confidence to communication overhead, while a margin-based safety rule prevents low signal-to-noise ratio (SNR) locks. Experiments on DeepSense-6G (train: scenarios 42 and 44; test:43) with a 21-beam discrete Fourier transform (DFT) codebook achieve Top-1/Top-3 of 0.81/0.99 with expected beam probe of 2 per sweep and zero observed outages at θ = 0 dB with margin Δ = 3 dB. The results show that multimodal priors with ensemble uncertainty match link quality and improve reliability compared to ablations while cutting overhead with better predictive model.

eess.SP↗

arXiv2026

Dual-Chirp AFDM for Joint Delay-Doppler Estimation with Rydberg Atomic Quantum Receivers

In this paper, we propose a joint delay-Doppler estimation framework for Rydberg atomic quantum receivers (RAQRs) leveraging affine frequency division multiplexing (AFDM), as a future enabler of hyper integrated sensing and communication (ISAC) in 6G and beyond. The proposed approach preserves the extreme sensitivity of RAQRs, while offering a pioneering solution to the joint estimation of delay-Doppler parameters of mobile targets, which has yet to be addressed in the literature due to the inherent coupling of time-frequency parameters in the optical readout of RAQRs to the best of our knowledge. To overcome this unavoidable ambiguity, we propose a dual-chirp AFDM framework where the utilization of distinct chirp parameters effectively converts the otherwise ambiguous estimation problem into a full-rank system, enabling unique delay-Doppler parameter extraction from RAQRs. Numerical simulations verify that the proposed dual-chirp AFDM shows superior delay-Doppler estimation performance compared to the classical single-chirp AFDM over RAQRs.

eess.SP↗

arXiv2026

Environment-aware Near-field UE Tracking under Partial Blockage and Reflection

This paper proposes an environment-aware near-field (NF) user equipment (UE) tracking method for extremely large aperture arrays. By integrating known surface geometries and tracking the line-of-sight (LOS) and non-line-of-sight (NLOS) indicators per antenna element, the method captures partial blockages and reflections specific to the NF spherical-wavefront regime, which are unavailable under the conventional far-field (FF) assumption. The UE positions are tracked by maximizing the cosine similarity between the predicted and received channels, enabling tracking even under complete LOS obstruction. Simulation results confirm that increasing environment-awareness improves accuracy, and that NF consistently outperforms FF baselines, achieving a $0.22\,\mathrm{m}$ root-mean-square error with full environment-awareness.

eess.SP↗

arXiv2026

Iterative Distillation for Reward-Guided Fine-Tuning of Diffusion Models in Biomolecular Design

We address the problem of fine-tuning diffusion models for reward-guided generation in biomolecular design. While diffusion models have proven highly effective in modeling complex, high-dimensional data distributions, real-world applications often demand more than high-fidelity generation, requiring optimization with respect to potentially non-differentiable reward functions such as physics-based simulation or rewards based on scientific knowledge. Although RL methods have been explored to fine-tune diffusion models for such objectives, they often suffer from instability, low sample efficiency, and mode collapse due to their on-policy nature. In this work, we propose an iterative distillation-based fine-tuning framework that enables diffusion models to optimize for arbitrary reward functions. Our method casts the problem as policy distillation: it collects off-policy data during the roll-in phase, simulates reward-based soft-optimal policies during roll-out, and updates the model by minimizing the KL divergence between the simulated soft-optimal policy and the current model policy. Our off-policy formulation, combined with KL divergence minimization, enhances training stability and sample efficiency compared to existing RL-based methods. Empirical results demonstrate the effectiveness and superior reward optimization of our approach across diverse tasks in protein, small molecule, and regulatory DNA design. The source code is released at (https://divelab.github.io/VIDD/).

cs.LG↗

arXiv2026

Personalized Parameter-Efficient Fine-Tuning of Foundation Models for Multimodal Recommendation

In recent years, substantial research has integrated multimodal item metadata into recommender systems, often by using pre-trained multimodal foundation models to encode such data. Since these models are not originally trained for recommendation tasks, recent works efficiently adapt them via parameter-efficient fine-tuning (PEFT). However, even with PEFT, item embeddings from multimodal foundation models remain user-blind: item embeddings are not conditioned on user interests, despite the fact that users with diverse interests attend to different item aspects. To address this limitation, we propose PerPEFT, a personalized PEFT strategy for multimodal recommendation. Specifically, PerPEFT groups users by interest and assigns a distinct PEFT module to each group, enabling each module to capture the fine-grained item aspects most predictive of that group`s purchase decisions. We further introduce a specialized training technique that strengthens this user-group conditioning. Notably, PerPEFT is PEFT-agnostic and can be paired with any PEFT method applicable to multimodal foundation models. Through extensive experiments, we show that (1) PerPEFT outperforms the strongest baseline by up to 15.3% (NDCG@20) and (2) delivers consistent gains across diverse PEFT variants. It is noteworthy that, even with personalization, PEFT remains lightweight, adding only 1.3% of the parameter count of the foundation model. We provide our code and datasets at https://github.com/kswoo97/PerPEFT.

cs.IR↗