arXiv subjects

Shreyas Kadekodi

Publications and source records attributed to Shreyas Kadekodi.

4 recordsLinked to original sources

arXiv2026

LeRoPE: Learnable RoPE Frequencies Improve Language Modeling

Rotary Positional Encodings (RoPE) are currently the most popular positional encodings used in modern language models. RoPE rotates two-dimensional chunks of query and key vectors, operating as a function of their relative positional offset. The position-wise rates of rotation in RoPE typically follow a geometric sequence specified by a fixed base-frequency hyperparameter. Prior work has improved performance by either increasing this parameter to slow rotation or by applying RoPE to only a subset of QK dimensions. In this work we modify RoPE by learning a scalar per frequency, treating frequencies as learnable parameters rather than hyperparameters. We validate Learned RoPE by training a ladder of language models from scratch, ranging from 52M to 2.5B parameters. We observe and analyze the emergence of a high-norm, positional LeRoPE band. LeRoPE consistently outperforms RoPE and partial RoPE across all scales, with RoPE requiring 3.4% more compute (FLOPs) to match LeRoPE at the largest scale.

cs.LG↗

arXiv2026

Measuring What Matters: Synthetic Benchmarks for Concept Bottleneck Models

Concept bottleneck models predict outcomes from high-level concepts detected in inputs. Although concepts provide a simple way to reap benefits from interpretability, very few datasets include concept labels. This limits researchers' ability to determine which problems are suitable for these models, isolate the factors that drive their performance or lead to failures, or uncover which algorithms perform well. In this paper, we develop synthetic benchmarks for concept-bottleneck models, focusing on their two main use cases: decision support, in which models assist humans in making better decisions, and automation, in which models handle routine tasks without supervision. Our benchmarks can generate labeled datasets while controlling for properties that affect performance, including data modality, concept choice, annotation quality, and completeness. We demonstrate how the benchmarks can be used to evaluate representative classes of concept bottleneck models. Our demonstrations show how the benchmarks can diagnose failure modes and guide follow-up testing.

cs.LG↗

arXiv2025

A Framework for the Systematic Evaluation of Obstacle Avoidance and Object-Aware Controllers

Real-time control is an essential aspect of safe robot operation in the real world with dynamic objects. We present a framework for the analysis of object-aware controllers, methods for altering a robot's motion to anticipate and avoid possible collisions. This framework is focused on three design considerations: kinematics, motion profiles, and virtual constraints. Additionally, the analysis in this work relies on verification of robot behaviors using fundamental robot-obstacle experimental scenarios. To showcase the effectiveness of our method we compare three representative object-aware controllers. The comparison uses metrics originating from the design considerations. From the analysis, we find that the design of object-aware controllers often lacks kinematic considerations, continuity of control points, and stability in movement profiles. We conclude that this framework can be used in the future to design, compare, and benchmark obstacle avoidance methods.

cs.RO↗

arXiv2025

CAT-RRT: Motion Planning that Admits Contact One Link at a Time

Current motion planning approaches rely on binary collision checking to evaluate the validity of a state and thereby dictate where the robot is allowed to move. This approach leaves little room for robots to engage in contact with an object, as is often necessary when operating in densely cluttered spaces. In this work, we propose an alternative method that considers contact states as high-cost states that the robot should avoid but can traverse if necessary to complete a task. More specifically, we introduce Contact Admissible Transition-based Rapidly exploring Random Trees (CAT-RRT), a planner that uses a novel per-link cost heuristic to find a path by traversing high-cost obstacle regions. Through extensive testing, we find that state-of-the-art optimization planners tend to over-explore low-cost states, which leads to slow and inefficient convergence to contact regions. Conversely, CAT-RRT searches both low and high-cost regions simultaneously with an adaptive thresholding mechanism carried out at each robot link. This leads to paths with a balance between efficiency, path length, and contact cost.

cs.RO↗