Recent advances in large language models (LLMs) enable agentic systems trained with reinforcement learning (RL) over multi-turn interaction trajectories, but practical deployment is bottlenecked by rapidly growing textual histories that inflate token budgets and memory usage. We introduce AgentOCR, a framework that exploits the superior information density of visual tokens by representing the accumulated observation-action history as a compact rendered image. To make multi-turn rollouts scalable, AgentOCR proposes segment optical caching. By decomposing history into hashable segments and maintaining a visual cache, this mechanism eliminates redundant re-rendering. Beyond fixed rendering, AgentOCR introduces agentic self-compression, where the agent actively emits a compression rate and is trained with compression-aware reward to adaptively balance task success and token efficiency. We conduct extensive experiments on challenging agentic benchmarks, ALFWorld and search-based QA. Remarkably, results demonstrate that AgentOCR preserves over 95\% of text-based agent performance while substantially reducing token consumption (>50\%), yielding consistent token and memory efficiency. Our further analysis validates a 20x rendering speedup from segment optical caching and the effective strategic balancing of self-compression.
Wikipedia's founders could not have dreamed they were creating the most important laboratory for social scientific and computing research in history but that is exactly what happened. Hill and Shaw take account of Wikipedia's enormous effect on academic scholarship
While large language models (LLMs) have proven effective in leveraging textual data for recommendations, their application to multimodal recommendation tasks remains relatively underexplored. Although LLMs can process multimodal information through projection functions that map visual features into their semantic space, recommendation tasks often require representing users' history interactions through lengthy prompts combining text and visual elements, which not only hampers training and inference efficiency but also makes it difficult for the model to accurately capture user preferences from complex and extended prompts, leading to reduced recommendation performance. To address this challenge, we introduce HistLLM, an innovative multimodal recommendation framework that integrates textual and visual features through a User History Encoding Module (UHEM), compressing multimodal user history interactions into a single token representation, effectively facilitating LLMs in processing user preferences. Extensive experiments demonstrate the effectiveness and efficiency of our proposed mechanism.
Keeping three radical ideas of Dr. B.R. Ambedkar, which have not been seriously dealt with by mainstream Indian/Kerala historiography, at the backdrop, namely, the Nagas and Dravidians are the same people, the untouchables were Buddhists, and India’s history as the history of mortal conflicts between Buddhism and Brahminism, the article attempts to study a Malayalam poem that has attained a classical status in the language, Poothapattu, to unravel the concealed layers of Kerala’s past. Drawing on the distinction the filmmaker Sergei Eisenstein establishes between the image and representation and on the insights provided by the Sangham Thinai conceptualizations, the article argues that in the Pootham image created by the Savarnna Malayalees, one could see sedimentation of history, where representations of the untouchable population of different historical moments are fused into a complex image, attesting to the veracity of Ambedkar’s radical ideas enumerated above.
Abinash Kumar Shaw, Raghunath Ghara, Paz Beniamini
et al.
We propose different estimators to probe the intergalactic medium (IGM) during epoch of reionization (EoR) using the dispersion measure (${\rm DM}$) of the fast radio bursts. We consider three different reionization histories, which we can distinguish with a total of $\lesssim 1000\,{\rm DM}$ measurements during EoR if their redshifts are known. We note that the redshift derivatives of ${\rm DM}$ are also directly sensitive to the reionization history. The major point of this work is to explore the variance in the ${\rm DM}$ measurements and the information encoded in them. We find that the all-sky average $\overline{\rm DM}(z)$ gets biased from the line-of-sight (LoS) fluctuations in the ${\rm DM}$ measurements introduced by the ionization of IGM during EoR. We find that the ratio $σ_{\rm DM}/\overline{\rm DM}$ depends directly on the ionization bubble sizes as well as the reionization history. On the other hand, we also find that angular variance (coined as $\textit{structure function}$) of ${\rm DM}$ encodes the information about the duration of reionization and the typical bubble sizes as well. We establish the usefulness of variances in ${\rm DM}$ using toy models of reionization and later verify it with the realistic reionization simulations.
Lekshmi Thulasidharan, Elena D'Onghia, Robert Benjamin
et al.
The prevailing model of galaxy formation proposes that galaxies like the Milky Way are built through a series of mergers with smaller galaxies over time. However, the exact details of the Milky Way's assembly history remain uncertain. In this study, we show that the Milky Way's merger history is uniquely encoded in the vertical thickness of its stellar disk. By leveraging age estimates from the value-added LAMOST DR8 catalog and the StarHorse ages from SDSS-IV DR12 data, we investigate the relationship between disk thickness and stellar ages in the Milky Way using a sample comprising Red Giants (RG), Red Clump Giants (RCG), and metal-poor stars (MPS). Guided by the IllustrisTNG50 simulations, we show that an increase in the dispersion of the vertical displacement of stars in the disk traces its merger history. This analysis reveals the epoch of a major merger event that assembled the Milky Way approximately 11.13 billion years ago, as indicated by the abrupt increase in disk thickness among stars of that age, likely corresponding to the Gaia-Sausage Enceladus (GSE) event. The data do not exclude an earlier major merger, which may have occurred about 1.3 billion years after the Big Bang. Furthermore, the analysis suggests that the geometric thick disk of the Milky Way was formed around 11.13 billion years ago, followed by a transition period of approximately 2.6 billion years leading to the formation of the geometric thin disk, illustrating the galaxy's structural evolution. Additionally, we identified three more recent events -- 5.20 billion, 2.02 billion, and 0.22 billion years ago -- potentially linked to multiple passages of the Sagittarius dwarf galaxy. Our study not only elucidates the complex mass assembly history of the Milky Way and highlights its past interactions but also introduces a refined method for examining the merger histories of external galaxies.