arXiv subjects

Sangwoo Kim

Publications and source records attributed to Sangwoo Kim.

8 recordsLinked to original sources

arXiv2026

How Do Language Models Represent and Use Phonological Information for Allomorph Selection?

Language models are trained on tokenized text that obscures the sound structure of words, yet they reliably produce morphemes whose form is phonologically conditioned. It remains unclear whether they rely on item-specific memorization or rule-like generalization and, if the latter, how that generalization is implemented. We therefore ask whether this phonological condition is represented within language models and how it is causally used for allomorph selection. For the English indefinite article a/an, we show that the phonological condition is encoded along a single linear direction in trigger-token embeddings, that this direction causally drives article selection in token-level wug tests, and that, at the article-prediction position, the model forecasts the upcoming trigger token and uses the forecasted trigger's phonological feature to choose the article. We then ask whether this rule-like generalization extends beyond English article selection, both to allomorph selection in other languages and to explicit phonological judgment. Together, these results provide a mechanistic account of phonologically conditioned allomorph selection in language models, and dissociate this generation-time ability from explicit metalinguistic judgments.

cs.CL↗

arXiv2026

Non-linear Interventions on Large Language Models

Intervention is one of the most representative and widely used methods for understanding the internal representations of large language models (LLMs). However, existing intervention methods are confined to linear interventions grounded in the Linear Representation Hypothesis, leaving features encoded along non-linear manifolds beyond their reach. In this work, we introduce a general formulation of intervention that extends naturally to non-linearly represented features, together with a learning procedure that further enables intervention on implicit features lacking a direct output signature. We validate our framework on refusal bypass steering, where it steers the model more precisely than linear baselines by intervening on a non-linear feature governing refusal.

cs.CL↗

arXiv2026

RLDX-1 Technical Report

While Vision-Language-Action models (VLAs) have shown remarkable progress toward human-like generalist robotic policies through the versatile intelligence (i.e. broad scene understanding and language-conditioned generalization) inherited from pre-trained Vision-Language Models, they still struggle with complex real-world tasks requiring broader functional capabilities (e.g. motion awareness, long-term memory, and physical sensing). To address this, we introduce RLDX-1, a general-purpose robotic policy for dexterous manipulation built on the Multi-Stream Action Transformer (MSAT), an architecture that unifies these capabilities by integrating heterogeneous modalities through modality-specific streams with cross-modal joint self-attention. RLDX-1 further combines this architecture with system-level design choices, including data synthesis for rare manipulation scenarios, learning procedures specialized for human-like manipulation, and inference optimizations for real-time deployment. Through empirical evaluation, we show that RLDX-1 consistently outperforms recent frontier VLAs (e.g. $\pi_{0.5}$ and GR00T N1.6) across both simulation benchmarks and real-world tasks that require broad functional capabilities beyond general versatility. In particular, RLDX-1 shows superiority in ALLEX humanoid tasks by achieving success rates of 86.8% while $\pi_{0.5}$ and GR00T N1.6 achieve around 40%, highlighting the ability of RLDX-1 to control a high-DoF humanoid robot under diverse functional demands. Together, these results position RLDX-1 as a promising step toward reliable VLAs for complex, contact-rich, and dynamic real-world dexterous manipulation.

cs.RO↗

arXiv2026

Universal Persistent Brownian Motions in Confluent Tissues

Biological tissues are active materials whose non-equilibrium dynamics emerge from distinct cellular force-generating mechanisms. Using a two-dimensional active foam model, we compare the effects of traction forces and junctional tension fluctuations on confluent tissue dynamics. While these two modes of activity produce qualitatively different cell shapes, rearrangement statistics, and spatiotemporal correlations in fluid states, we find that the long-time cellular motion universally converges to persistent Brownian dynamics. This universal feature contrasts with the non-universal correlations between cell geometry, rearrangement rate, and fluidity, which depend sensitively on the underlying modes of active force. Our results demonstrate that persistent Brownian motion provides a minimal framework for describing tissue dynamics, while distinct active forces leave identifiable structural and dynamical signatures, thereby enabling inference of the dominant active force in fluid state tissues.

physics.bio-ph↗

arXiv2025

SUIT: Knowledge Editing with Subspace-Aware Key-Value Mappings

Knowledge editing aims to efficiently correct factual errors in language models. Widely used locate-then-edit methods update an MLP layer by adjusting its weights to change the mapping between the layer's input vector (key) and output vector (value), thereby editing the model's knowledge. As this update is driven by key and value vectors, obtaining these vectors without careful constraints causes significant model perturbations beyond the targeted edit, a common issue in many prior knowledge editing methods. To address this, we propose Subspace Knowledge Edit (SUIT), which computes key and value vectors only within the subspace of critical features relevant to the edit. Our empirical results on LLaMA3, GPT-J, and Qwen2.5 models show that SUIT dramatically improves knowledge preservation over strong baselines while maintaining high editing performance. These results support the claim that SUIT successfully identifies the critical subspace for the edit. Beyond quantitative gains, our analyses show that SUIT reduces unintended perturbations in hidden states while confining updates to directions that are more effective for editing. Taken together, these findings establish edit-critical subspace identification as a key principle for reliable, low-perturbation knowledge editing. Our code is available at https://github.com/holi-lab/SUIT.

cs.CL↗

arXiv2025

Saliency-Aware Quantized Imitation Learning for Efficient Robotic Control

Deep neural network (DNN)-based policy models, such as vision-language-action (VLA) models, excel at automating complex decision-making from multi-modal inputs. However, scaling these models greatly increases computational overhead, complicating deployment in resource-constrained settings like robot manipulation and autonomous driving. To address this, we propose Saliency-Aware Quantized Imitation Learning (SQIL), which combines quantization-aware training with a selective loss-weighting strategy for mission-critical states. By identifying these states via saliency scores and emphasizing them in the training loss, SQIL preserves decision fidelity under low-bit precision. We validate SQIL's generalization capability across extensive simulation benchmarks with environment variations, real-world tasks, and cross-domain tasks (self-driving, physics simulation), consistently recovering full-precision performance. Notably, a 4-bit weight-quantized VLA model for robotic manipulation achieves up to 2.5x speedup and 2.5x energy savings on an edge GPU with minimal accuracy loss. These results underline SQIL's potential for efficiently deploying large IL-based policy models on resource-limited devices.

cs.RO↗

arXiv2025

Restricted Gravity: Lagrangian Formalism and Explicit Solutions

It is well known that, making the Abelian projection of Einstein's theory one can obtain the restricted gravity which is simpler than Einstein's theory but describes the core dynamics of Einstein's gravity. In this paper we present the Lagrangian formalism of the restricted gravity which makes the restricted gravity a self consistent field theory by itself, independent of Einstein's theory. With this we present interesting solutions of the restricted gravity, in particular the gravitational cosmic string, the Bertotti-Robinson spacetime, the axisymmetric pp-wave, and the conformally flat waves with flat wavefront. Moreover, we show that in the restricted gravity the Rosen-Bondi gravitational plane wave could be described by two spin-one Maxwellian potentials. This could play important role in quantum gravity. We discuss the physical implications of the restricted gravity.

gr-qc↗

arXiv2024

Exceptionally Dense and Resilient Polydisperse Disk Packings

Understanding the way disordered particle packings transition between jammed (rigid) and unjammed (fluid) states is of both great practical importance and strong fundamental interest. The values of critical packing fraction (and other state variables) at the jamming transition are protocol dependent. Here, we demonstrate that this variability can be systematically traced to structural measures of packing, as well as to energy measures inside the jamming regime. A novel generalized simultaneous particle swap algorithm constructs overjammed states of desired energy, which upon decompression lead to predictable critical packing fractions. Thus, for a given set of particle sizes, states with extraordinarily high critical packing fractions can be found efficiently, which sustain substantial shear strain and preserve their special structure over the entire jammed domain. The close relation revealed here between the energy landscape of overjammed soft-particle packings and the behavior near the jamming transition points towards new ways of understanding and constructing disordered materials with exceptional properties.

cond-mat.soft↗