arXiv subjects

Saransh Agrawal

Publications and source records attributed to Saransh Agrawal.

3 recordsLinked to original sources

arXiv2026

Adaptive Helpfulness-Harmlessness Alignment with Preference Vectors

Ensuring that large language models (LLMs) are both helpful and harmless is a critical challenge, as overly strict constraints can lead to excessive refusals, while permissive models risk generating harmful content. Existing approaches, such as reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO), attempt to balance these trade-offs but suffer from performance conflicts, limited controllability, and poor extendability. To address these issues, we propose Preference Vector, a novel framework inspired by task arithmetic. Instead of optimizing multiple preferences within a single objective, we train separate models on individual preferences, extract behavior shifts as preference vectors, and dynamically merge them at test time. This modular approach enables fine-grained, user-controllable preference adjustments and facilitates seamless integration of new preferences without retraining. Experiments show that our proposed Preference Vector framework improves helpfulness without excessive conservatism, allows smooth control over preference trade-offs, and supports scalable multi-preference alignment.

cs.LG↗

arXiv2026

Spiderbot: An Open-Source Energy-Efficient Hexapod with Passive Gravity Compensation

Hexapod robots can achieve static stability with fewer actuated joints than bipeds or quadrupeds, yet many platforms still use 3-DOF legs, increasing weight and continuous torque requirement with limited gain in locomotion capability on flat, inclined and moderately rough terrains. We release Spiderbot, an open-source hexapod that uses a 4-bar linkage with a passive spring to mechanically support body weight, with a 2-DOF per-leg design that substantially reduces energy consumption. This mechanism substantially offloads gravitational torque during standing stance consuming only 1.5W (reduction of over 90\% over the unsprung version and up to 96\% over other similar hexapods). The passive spring compensation extends to payloads of up to 3.25kg with no additional torque requirements. The platform enables long-duration deployments on a modest battery budget and costs under \$400, making it suitable for large-scale multi-agent experiments. We validate the locomotion capabilities of the platform with an RL policy trained in mjlab, including successful sim-to-real transfer, despite the complexity of the mechanism. The platform is evaluated on flat and rough terrains, slope up to $15^\circ$ and step obstacles. We release all the CAD files, assembling instructions, and full training and deployment code along with the model checkpoints at https://erc-bpgc.github.io/SpiderBot/.

cs.RO↗

arXiv2025

SHA256 at SemEval-2025 Task 4: Selective Amnesia -- Constrained Unlearning for Large Language Models via Knowledge Isolation

Large language models (LLMs) frequently memorize sensitive information during training, posing risks when deploying publicly accessible models. Current machine unlearning methods struggle to selectively remove specific data associations without degrading overall model capabilities. This paper presents our solution to SemEval-2025 Task 4 on targeted unlearning, which introduces a two-stage methodology that combines causal mediation analysis with layer-specific optimization. Through systematic causal tracing experiments on OLMo architectures (1B and 7B parameters), we identify the critical role of the first few transformer layers (layers 0-5) in storing subject-attribute associations within MLP modules. Building on this insight, we develop a constrained optimization approach that freezes upper layers while applying a novel joint loss function to lower layers-simultaneously maximizing forget set loss via output token cross-entropy penalties and minimizing retain set deviation through adaptive regularization. Our method achieves 2nd place in the 1B model track, demonstrating strong task performance while maintaining 88% of baseline MMLU accuracy. These results establish causal-informed layer optimization as a promising paradigm for efficient, precise unlearning in LLMs, offering a significant step forward in addressing data privacy concerns in AI systems.

cs.CL↗