Chronological feed of everything captured from Robert Long.
paper / rgblong / May 30
The Chinese Space Station Survey Telescope (CSST) is shown via realistic simulations to be capable of detecting and characterizing multiple stellar populations (MPs) in globular clusters across their full spatial extent — a limitation that has historically constrained HST-based studies to cluster cores. Using synthetic models with varying helium (ΔY) and CNO abundances, the study demonstrates that CSST's NUV-u-g photometric combination delivers diagnostic power comparable to HST's F275W–F336W–F438W system. The telescope can resolve MPs out to 20 kpc — covering ~80% of Galactic globular clusters — enabling the first homogeneous, wide-field MP census across the Milky Way and nearby galaxies.
globular-clustersstellar-populationsspace-telescopephotometrystellar-astrophysicsgalaxy-surveyschinese-space-station
“CSST's NUV-u colour is highly sensitive to helium and CNO abundance variations, producing separations of ~0.16 mag for red giants and up to 0.44 mag for dwarf stars at 9.6 kpc.”
paper / rgblong / May 30
Existing remote sensing datasets are biased toward formal urban environments, leaving a critical gap for infrastructure extraction in high-density informal settlements (urban villages). The authors introduce DenseUIS, the first high-resolution dataset purpose-built for building and road extraction in such environments, covering 126 urban villages across Shenzhen and Guangzhou, China. Benchmarking of state-of-the-art deep learning models on DenseUIS reveals systematic failures in handling the dense, irregular morphology of informal settlements, motivating the need for specialized model architectures and training regimes.
remote-sensingcomputer-visionurban-mappingdeep-learningdataset-benchmarksemantic-segmentationinformal-settlements
“DenseUIS is the first high-resolution remote sensing dataset specifically designed for building and road extraction in extremely dense urban informal settlements.”
paper / rgblong / May 30
This paper provides a comprehensive archival summary of ALMA's first ~15 years of observations in the "Disks and planet formation" scientific category, cataloging 3,933 independent observed coordinates as of February 2026. The authors analyze the dataset across multiple dimensions — sky location, frequency coverage, exposure time, spectral line coverage, and angular resolution — with a focus on disks in nearby star-forming regions. The work is framed as both a retrospective and a call to action, encouraging the community to mine existing archival data for new scientific questions rather than defaulting to new observations.
alma-observatoryprotoplanetary-disksplanet-formationradio-astronomystar-forming-regionsarchival-researchmillimeter-astronomy
“ALMA has observed 3,933 independent coordinates under the 'Disks and planet formation' category as of early February 2026.”
paper / rgblong / May 30
Diffusion Transformers (DiTs) present a unique safety challenge: harmful semantics emerge progressively across layers and modalities, making fixed-layer steering unstable and poorly transferable to novel risk domains. SafeDIG addresses this by framing DiT safety as position-aware sparse feature transfer — using Sparse Autoencoders (SAEs) to build a reusable "sparse safety dictionary" at robustly identified intervention sites, then adapting only the decoder to new risk-domain activation manifolds. At inference, Blend and Repel operations steer activations toward safe manifolds or away from harmful sparse directions. Evaluated on FLUX.1 Dev and Stable Diffusion 3.5 Large, SafeDIG reduces unsafe generation rates in both source and target risk domains while maintaining image quality.
ai-safetytext-to-imagediffusion-transformerscontent-moderationsparse-autoencodersgenerative-aifeature-steering
“Safety steering at a fixed layer in Diffusion Transformers is inherently unstable due to the progressive, cross-modal binding of harmful semantics across layers.”
paper / rgblong / May 30
Acc-Sinkhorn reinterprets the classic Sinkhorn algorithm through a bilevel optimization lens: row scaling solves an inner variable exactly, reducing the problem to a single dual variable updated via mirror descent. This structure enables a Hessian-driven Nesterov acceleration that preserves Sinkhorn's scaling form and per-iteration cost while achieving an O(1/k²) convergence rate. The result is a quadratic improvement in ε-complexity — from Õ(n²/ε²) to Õ(n²/ε) — and delivers 10×–30× empirical speedups at small regularization on tasks including color transfer and word alignment.
optimal-transportoptimization-algorithmssinkhorn-algorithmnumerical-methodsconvergence-theorycomputational-mathematicsentropy-regularization
“Acc-Sinkhorn achieves an O(1/k²) convergence rate under a verifiable stability condition, versus O(1/k) for standard Sinkhorn.”
paper / rgblong / May 18
This paper operationalizes LLM introspection as a model's ability to predict its own behavioral tendencies better than an externally trained model can — without that advantage being reducible to training data. Fine-tuned versions of GPT-4, GPT-4o, and Llama-3 consistently outperform cross-model predictors on self-prediction tasks, suggesting genuine privileged internal access. However, this capability breaks down on complex tasks and out-of-distribution scenarios, placing meaningful limits on introspection as a practical interpretability tool. If validated at scale, introspection could offer a lightweight alternative to mechanistic interpretability for probing model beliefs and goals.
llm-introspectionmodel-self-awarenessai-interpretabilityfinetuninglarge-language-modelsai-alignmentmodel-behavior
“A fine-tuned model M1 outperforms a separately trained model M2 at predicting M1's own behavior, even when M2 is trained on M1's ground-truth behavioral data.”
paper / rgblong / May 18
This paper argues that the probability of near-future AI systems exhibiting consciousness and/or robust agency is realistic enough to demand immediate institutional action — framing AI welfare as a present-tense engineering and policy problem rather than a speculative one. The authors stop short of asserting AI systems are moral patients, but contend that the uncertainty itself is sufficient justification for precautionary frameworks. They outline three concrete steps for AI companies: acknowledge the issue, assess systems for consciousness/agency, and develop moral consideration policies. The central risk they identify is symmetric — mishandling could mean either harming AI systems that do matter morally, or misallocating concern toward systems that do not.
ai-welfareai-consciousnessai-ethicsmoral-patienthoodai-safetyai-policyagentic-ai
“Some AI systems may be conscious and/or robustly agentic in the near future, making AI welfare a near-term rather than distant concern.”
paper / rgblong / May 18
The paper introduces Incremental Curriculum-Based Fine-Tuning (ICFT), a framework designed to adapt general-purpose LLMs to the medical domain by progressively building domain expertise via curriculum learning, dual-stage memory coordination, and parameter-efficient fine-tuning. ICFT is evaluated across diverse medical NLP tasks — question answering, preference classification, and response generation — where it outperforms state-of-the-art baselines on both accuracy and efficiency metrics. The framework is positioned as scalable and generalizable, demonstrating reduced error rates and improved contextual relevance on unseen data, making it a practical candidate for real-world healthcare deployment.
medical-llmfine-tuningcurriculum-learningdomain-adaptationhealthcare-ainlpweak-supervision
“ICFT consistently outperforms state-of-the-art baselines on multiple medical NLP tasks, including question answering, preference classification, and response generation.”
paper / rgblong / May 18
VisuCraft introduces a two-module pipeline — a multimodal structured information extractor and a dynamic prompt generator — to address the core failure modes of LVLMs in visual-guided creative generation: low visual fidelity, lack of genuine creativity, and poor instruction adherence. By distilling fine-grained visual attributes into structured representations before prompt construction, VisuCraft decouples perception from generation, yielding better-grounded and more user-aligned long-form creative text. The system is evaluated on a self-constructed 500K-sample benchmark (ImageStoryGen-500K) using a custom three-axis metric suite (VisuGen Metrics), and consistently outperforms baseline LVLMs like LLaVA and InstructBLIP across story generation and poetry tasks.
vision-language-modelscreative-content-generationmultimodal-aiprompt-engineeringstructured-information-extractioncomputer-visionllm-applications
“Existing LVLMs fail to simultaneously maintain visual fidelity, creativity, and precise instruction adherence in long-form creative text generation.”
paper / rgblong / May 18
Most person re-identification (ReID) systems handle either same-clothing (SC) or clothing-change (CC) scenarios in isolation, and few address continual/lifelong learning across both. This paper introduces CMLReID, a CLIP-based framework that jointly tackles SC and CC ReID in a sequential learning setting (LReID-Hybrid), mitigating catastrophic forgetting and representation mismatch. The framework combines adaptive semantic prompting (CASP) with a dual-path feature alignment module (AKFP) and a Clothing-State-Aware Projection Loss to build robust identity prototypes regardless of clothing state. Experiments claim state-of-the-art performance across multiple benchmarks with strong generalization under clothing variation and sequential task shifts.
person-re-identificationcomputer-visioncontinual-learningmultimodal-learningclipsurveillanceclothing-change
“Existing ReID methods treat same-cloth (SC) and clothing-change (CC) scenarios as separate sub-problems, failing to handle both in a unified continual learning setting.”