arXiv subjects

Sitaram Asur

Publications and source records attributed to Sitaram Asur.

14 recordsLinked to original sources

arXiv2026

Co-Evolving Harnesses and Models: On-Policy Correction Helps Weaker Models Catch Up Where Imitation Fails

Agent harnesses (the system prompt, tool set, execution hooks, and context-management scaffolding around a model) are a critical determinant of agentic task success. Automated harness evolution can enable smaller models to perform well on domain-specific tasks at a fraction of frontier-model cost. Since both the harness and model weights shape behavior, we ask how harness evolution and lightweight fine-tuning should be combined. Across seven enterprise agent tasks, we first evolve a harness with the weaker model, then find that a stronger expert often uses it more effectively, suggesting expert supervision could close the remaining gap. However, training the weaker model on the expert's complete trajectories under the evolved harness backfires: performance regresses on all seven tasks by 4 to 30 points across Qwen3-Coder and Gemma 4, even though the same procedure helps under the unevolved harness. Our analysis shows that imitation transfers knowledge and increases scaffold usage, but disrupts model-harness fit: the weaker model adopts the expert's planning strategy without the competence to execute it and no longer matches the harness evolved around its native planning style. We therefore develop an on-policy expert-correction pipeline, automated by a meta-level MLE agent, that localizes the failing turn in the weaker model's own rollout and asks the expert to rewrite only that turn. This preserves the model's planning style and combines the gains of harness evolution and model adaptation. Our results identify and resolve a source of contention between harness and weight updates, yielding a compatibility-preserving recipe for economical co-evolution on domain-specific enterprise tasks.

cs.AI↗

arXiv2025

Diversity Enhances an LLM's Performance in RAG and Long-context Task

The rapid advancements in large language models (LLMs) have highlighted the challenge of context window limitations, primarily due to the quadratic time complexity of the self-attention mechanism (\(O(N^2)\), where \(N\) denotes the context window length). This constraint impacts tasks such as retrieval-augmented generation (RAG) in question answering (Q\&A) and long context summarization. A common approach involves selecting content with the highest similarity to the query; however, this often leads to redundancy and the exclusion of diverse yet relevant information. Building on principles from Maximal Marginal Relevance (MMR) and Farthest Point Sampling (FPS), we integrate diversity into the content selection process. Our findings reveal that incorporating diversity substantially increases the recall of selecting relevant sentences or chunks before LLM-based Q\&A and summarization. These results highlight the importance of maintaining diversity in future LLM applications to further improve summarization and Q\&A outcomes.

cs.CL↗

arXiv2024

Rate, Explain and Cite (REC): Enhanced Explanation and Attribution in Automatic Evaluation by Large Language Models

LLMs have demonstrated impressive proficiency in generating coherent and high-quality text, making them valuable across a range of text-generation tasks. However, rigorous evaluation of this generated content is crucial, as ensuring its quality remains a significant challenge due to persistent issues such as factual inaccuracies and hallucination. This paper introduces three fine-tuned general-purpose LLM autoevaluators, REC-8B, REC-12B and REC-70B, specifically designed to evaluate generated text across several dimensions: faithfulness, instruction following, coherence, and completeness. These models not only provide ratings for these metrics but also offer detailed explanation and verifiable citation, thereby enhancing trust in the content. Moreover, the models support various citation modes, accommodating different requirements for latency and granularity. Extensive evaluations on diverse benchmarks demonstrate that our general-purpose LLM auto-evaluator, REC-70B, outperforms state-of-the-art LLMs, excelling in content evaluation by delivering better quality explanation and citation with minimal bias. Our REC dataset and models are available at https://github.com/adelaidehsu/REC.

cs.CL↗

arXiv2024

UNA: A Unified Supervised Framework for Efficient LLM Alignment Across Feedback Types

RL alignment methods, including RLHF and DPO, are primarily based on pairwise preference data. Although scalar or score-based feedback has been collected in some settings, it is rarely used directly, and preference magnitude information is typically ignored. Furthermore, current alignment frameworks offer limited capability for unifying heterogeneous supervision signals, making it difficult to jointly leverage diverse data types within a single training paradigm. This limitation constrains the richness and scalability of the alignment process. To address this gap, we propose a \textbf{UN}ified \textbf{A}lignment (UNA) framework capable of training across different types of feedback, including binary, pairwise, and score-based, through a generalized implicit reward function. The reward function is theoretically proved to be the optimal policy by the log sum inequality. Extensive experiments on classical benchmarks consistently demonstrate the advantage of the proposed unified framework with typical LLM base models.

cs.LG↗

arXiv2024

Reinforcement Learning for LLM Post-Training: A Survey

Large language models (LLMs) trained via pretraining and supervised fine-tuning (SFT) can still produce harmful and misaligned outputs, or struggle in domains like math and coding. Reinforcement learning (RL)-based post-training methods, including Reinforcement Learning from Human Feedback (RLHF) methods like Direct Preference Optimization (DPO) and Reinforcement Learning with Verifiable Rewards (RLVR) approaches like PPO and GRPO, have made remarkable gains to alleviate these issues. Yet, no existing work offers a technically detailed comparison of the various methods driving this progress. In order to fill this gap, we present a timely survey that connects foundational components with latest advancements. We derive a single policy gradient framework that unifies pretraining, SFT, RLHF, and RLVR as special cases while also organizing the more recent techniques therein. The main contributions of our survey are as follows: (1) a self-contained introduction to MLE, RLHF, and RLVR foundations and the unified policy gradient framework; (2) detailed technical analysis of PPO- and GRPO-based methods alongside offline and iterative DPO approaches, decomposed along prompt sampling, response sampling, and gradient coefficient axes; (3) standardized notation enabling direct cross-method comparison; and (4) comprehensive comparison of implementation details and empirical results of each method in the appendix. We aim to serve as a technically grounded reference for researchers and practitioners working on LLM post-training.

cs.CL↗

arXiv2024

PAFT: A Parallel Training Paradigm for Effective LLM Fine-Tuning

Large language models (LLMs) have shown remarkable abilities in diverse natural language processing (NLP) tasks. The LLMs generally undergo supervised fine-tuning (SFT) followed by preference alignment to be usable in downstream applications. However, this sequential training pipeline leads to alignment tax that degrades the LLM performance. This paper introduces PAFT, a new PArallel training paradigm for effective LLM Fine-Tuning, which independently performs SFT and preference alignment (e.g., DPO and ORPO, etc.) with the same pre-trained model on respective datasets. The model produced by SFT and the model from preference alignment are then merged into a final model by parameter fusing for use in downstream applications. This work reveals important findings that preference alignment like DPO naturally results in a sparse model while SFT leads to a natural dense model which needs to be sparsified for effective model merging. This paper introduces an effective interference resolution which reduces the redundancy by sparsifying the delta parameters. The LLM resulted from the new training paradigm achieved Rank #1 on the HuggingFace Open LLM Leaderboard. Comprehensive evaluation shows the effectiveness of the parallel training paradigm.

cs.CL↗

arXiv2014

Deciding what to display: maximizing the information value of social media

In information-rich environments, the competition for users' attention leads to a flood of content from which people often find hard to sort out the most relevant and useful pieces. Using Twitter as a case study, we applied an attention economy solution to generate the most informative tweets for its users. By considering the novelty and popularity of tweets as objective measures of their relevance and utility, we used the Huberman-Wu algorithm to automatically select the ones that will receive the most attention in the next time interval. Their predicted popularity was confirmed by using Twitter data collected for a period of 2 months.

cs.CY↗

arXiv2013

Dynamics of Trends and Attention in Chinese Social Media

There has been a tremendous rise in the growth of online social networks all over the world in recent years. It has facilitated users to generate a large amount of real-time content at an incessant rate, all competing with each other to attract enough attention and become popular trends. While Western online social networks such as Twitter have been well studied, the popular Chinese microblogging network Sina Weibo has had relatively lower exposure. In this paper, we analyze in detail the temporal aspect of trends and trend-setters in Sina Weibo, contrasting it with earlier observations in Twitter. We find that there is a vast difference in the content shared in China when compared to a global social network such as Twitter. In China, the trends are created almost entirely due to the retweets of media content such as jokes, images and videos, unlike Twitter where it has been shown that the trends tend to have more to do with current global events and news stories. We take a detailed look at the formation, persistence and decay of trends and examine the key topics that trend in Sina Weibo. One of our key findings is that retweets are much more common in Sina Weibo and contribute a lot to creating trends. When we look closer, we observe that most trends in Sina Weibo are due to the continuous retweets of a small percentage of fraudulent accounts. These fake accounts are set up to artificially inflate certain posts, causing them to shoot up into Sina Weibo's trending list, which are in turn displayed as the most popular topics to users.

cs.SI↗

arXiv2012

The Pulse of News in Social Media: Forecasting Popularity

News articles are extremely time sensitive by nature. There is also intense competition among news items to propagate as widely as possible. Hence, the task of predicting the popularity of news items on the social web is both interesting and challenging. Prior research has dealt with predicting eventual online popularity based on early popularity. It is most desirable, however, to predict the popularity of items prior to their release, fostering the possibility of appropriate decision making to modify an article and the manner of its publication. In this paper, we construct a multi-dimensional feature space derived from properties of an article and evaluate the efficacy of these features to serve as predictors of online popularity. We examine both regression and classification algorithms and demonstrate that despite randomness in human behavior, it is possible to predict ranges of popularity on twitter with an overall 84% accuracy. Our study also serves to illustrate the differences between traditionally prominent sources and those immensely popular on the social web.

cs.CY↗

arXiv2012

Artificial Inflation: The True Story of Trends in Sina Weibo

There has been a tremendous rise in the growth of online social networks all over the world in recent years. This has facilitated users to generate a large amount of real-time content at an incessant rate, all competing with each other to attract enough attention and become trends. While Western online social networks such as Twitter have been well studied, characteristics of the popular Chinese microblogging network Sina Weibo have not been. In this paper, we analyze in detail the temporal aspect of trends and trend-setters in Sina Weibo, constrasting it with earlier observations on Twitter. First, we look at the formation, persistence and decay of trends and examine the key topics that trend in Sina Weibo. One of our key findings is that retweets are much more common in Sina Weibo and contribute a lot to creating trends. When we look closer, we observe that a large percentage of trends in Sina Weibo are due to the continuous retweets of a small amount of fraudulent accounts. These fake accounts are set up to artificially inflate certain posts causing them to shoot up into Sina Weibo's trending list, which are in turn displayed as the most popular topics to users.

cs.CY↗

arXiv2011

What Trends in Chinese Social Media

There has been a tremendous rise in the growth of online social networks all over the world in recent times. While some networks like Twitter and Facebook have been well documented, the popular Chinese microblogging social network Sina Weibo has not been studied. In this work, we examine the key topics that trend on Sina Weibo and contrast them with our observations on Twitter. We find that there is a vast difference in the content shared in China, when compared to a global social network such as Twitter. In China, the trends are created almost entirely due to retweets of media content such as jokes, images and videos, whereas on Twitter, the trends tend to have more to do with current global events and news stories.

cs.CY↗

arXiv2011

Trends in Social Media : Persistence and Decay

Social media generates a prodigious wealth of real-time content at an incessant rate. From all the content that people create and share, only a few topics manage to attract enough attention to rise to the top and become temporal trends which are displayed to users. The question of what factors cause the formation and persistence of trends is an important one that has not been answered yet. In this paper, we conduct an intensive study of trending topics on Twitter and provide a theoretical basis for the formation, persistence and decay of trends. We also demonstrate empirically how factors such as user activity and number of followers do not contribute strongly to trend creation and its propagation. In fact, we find that the resonance of the content with the users of the social network plays a major role in causing trends.

cs.CY↗

arXiv2010

Influence and Passivity in Social Media

The ever-increasing amount of information flowing through Social Media forces the members of these networks to compete for attention and influence by relying on other people to spread their message. A large study of information propagation within Twitter reveals that the majority of users act as passive information consumers and do not forward the content to the network. Therefore, in order for individuals to become influential they must not only obtain attention and thus be popular, but also overcome user passivity. We propose an algorithm that determines the influence and passivity of users based on their information forwarding activity. An evaluation performed with a 2.5 million user dataset shows that our influence measure is a good predictor of URL clicks, outperforming several other measures that do not explicitly take user passivity into account. We also explicitly demonstrate that high popularity does not necessarily imply high influence and vice-versa.

cs.CY↗

arXiv2010

Predicting the Future with Social Media

In recent years, social media has become ubiquitous and important for social networking and content sharing. And yet, the content that is generated from these websites remains largely untapped. In this paper, we demonstrate how social media content can be used to predict real-world outcomes. In particular, we use the chatter from Twitter.com to forecast box-office revenues for movies. We show that a simple model built from the rate at which tweets are created about particular topics can outperform market-based predictors. We further demonstrate how sentiments extracted from Twitter can be further utilized to improve the forecasting power of social media.

cs.CY↗