Siddharth Yadav

Suraj Yadav, Siddharth Yadav, Parth Goyal

Recent alignment work on Large Language Models (LLMs) suggests preference optimization can improve reasoning by shifting probability mass toward better solutions. We test this claim in a resource-constrained setting by applying GRPO with LoRA to SLMs (up to 3B) for math reasoning on GSM8K and MATH datasets with difficulty-stratified analyses. As problem difficulty increases, accuracy plateaus, revealing a capacity boundary: GRPO primarily reshapes output preferences without reliably improving hardest-tier solving. Consistent with this, training GRPO only on lower-difficulty problems matches full-dataset accuracy across difficulty tiers while using only ~45% training steps, indicating diminishing returns from harder samples in this regime. We also find a cross-dataset generalization effect: GSM8K-trained GRPO achieves higher accuracy on the numeric subset of MATH than MATH-trained GRPO, exceeding it by ~5% at 1.5B and by ~3% at 3B. We show that the best achievable gains depend strongly on the base model's prior reasoning competence and the dataset's difficulty profile.

Khashayar Etemadi, Niloofar Tarighat, Siddharth Yadav et al.

The most natural method for evaluating program repair systems is to run them on bug datasets, such as Defects4J. Yet, using this evaluation technique on arbitrary real-world programs requires heavy configuration. In this paper, we propose a purely static method to evaluate the potential of the search space of repair approaches. This new method enables researchers and practitioners to encode the search spaces of repair approaches and select potentially useful ones without struggling with tool configuration and execution. We encode the search spaces by specifying the repair strategies they employ. Next, we use the specifications to check whether past commits lie in repair search spaces. For a repair approach, including many human-written past commits in its search space indicates its potential to generate useful patches. We implement our evaluation method in LighteR. LighteR gets a Git repository and outputs a list of commits whose source code changes lie in repair search spaces. We run LighteR on 55,309 commits from the history of 72 Github repositories with and show that LighteR's precision and recall are 77% and 92%, respectively. Overall, our experiments show that our novel method is both lightweight and effective to study the search space of program repair approaches.

Siddharth Yadav, Tanmoy Chakraborty

Code-mixing is the practice of alternating between two or more languages. Mostly observed in multilingual societies, its occurrence is increasing and therefore its importance. A major part of sentiment analysis research has been monolingual, and most of them perform poorly on code-mixed text. In this work, we introduce methods that use different kinds of multilingual and cross-lingual embeddings to efficiently transfer knowledge from monolingual text to code-mixed text for sentiment analysis of code-mixed text. Our methods can handle code-mixed text through a zero-shot learning. Our methods beat state-of-the-art on English-Spanish code-mixed sentiment analysis by absolute 3\% F1-score. We are able to achieve 0.58 F1-score (without parallel corpus) and 0.62 F1-score (with parallel corpus) on the same benchmark in a zero-shot way as compared to 0.68 F1-score in supervised settings. Our code is publicly available.