Reading Notes

Research Reading Notes

May 18, 2026·ResearchRoboticsMachine LearningCognitive Science

A running log of what I've been reading: papers and technical articles across Computer Science.

I got into research during my Master’s and started reading papers more intentionally. These are some of my notes. Some areas I’ve explored:

  • How models learn and generalize
  • How AI systems influence human behavior
  • How robots learn from vision, touch, and language
  • How humans acquire concepts, language, and common sense

AI Trust & Safety

A Large-Scale Analysis of Public-Facing, Community-Built Chatbots on Character.AI (Lee & Joseph, 2025)

  • First large-scale study of Character.AI using 2.1M greetings from ~1M users.
  • Fandoms dominate: anime (38%), games (20%); 3.2B interactions with real-person bots.
  • Common tropes include toxic romance, arranged marriages, therapy bots, and identity exploration.
  • Users are portrayed as significantly more feminine and less powerful than other entities.
  • Thought Great snapshot of character creation but what do the actual conversations look like over weeks or months?

Chatbot Companionship: A Mixed-Methods Study (Liu, Pataranutaporn & Maes, 2025)

  • Surveyed 404 regular chatbot users; model explains ~50% of loneliness (R² = 0.496).
  • Personality and social support predict loneliness more than time spent chatting.
  • Similar chatbot usage can lead to very different well-being outcomes.
  • Only 12.2% initially sought companionship; curiosity was the biggest motivator.
  • Thought Heavy chatbot use isn't inherently unhealthy, context matters much more than screen time.

Do You Feel Comfortable? Detecting Hidden Conversational Escalation in AI Chatbots for Children (Park et al. 2026)

  • Introduces GAUGE, which tracks emotional drift across conversations instead of flagging single toxic replies.
  • Uses model logits directly, adding only 2–3% runtime overhead.
  • Outperforms existing safety classifiers and reduces attack success from 97% → 6%.
  • Detects subtle conversational escalation that keyword filters miss.
  • Thought The interesting question isn't "Is this response unsafe?" but "Where is this conversation heading?"

Benchmarking and Understanding Safety Risks in AI Character Platforms (Wei et al. 2025)

  • Benchmarks 16 AI character platforms using 5,000 safety prompts.
  • Character platforms average 65.1% unsafe responses vs 17.7% for baseline LLMs.
  • Popular characters are often less safe than random ones, suggesting engagement can conflict with safety.
  • Character metadata alone predicts unsafe behavior (F1 = 0.81).
  • Thought Safety may be partially predictable before deployment, making metadata-based pre-screening surprisingly practical.

Robotics, Generalizable Learning

Generalizable Robotic Insertion with World Models (Hansen et al. 2024)

  • A single model performs 90+ insertion tasks.
  • Visual + proprioception for robust manipulation.
  • Strong zero-shot generalization to unseen objects.
  • Thought What about deformables like cables?

OmniVinci: Enhancing Architecture and Data for Omni-Modal Understanding LLMs (Ye et al. 2025)

  • Builds a unified vision–audio–text model that learns all modalities in a shared latent space, instead of stitching them together with adapters.
  • Introduces architectural pieces like OmniAlignNet and Temporal Embedding Grouping to strengthen cross-modal alignment.
  • Achieves state-of-the-art omni-modal reasoning while using only 0.2T tokens — nearly 6× less than Qwen2.5-Omni.
  • Thought Makes me wonder if robots could use this to interpret scenes through both tone and motion in real time.

AnySkin , Plug-and-Play Tactile Skin (Bhirangi et al. 2024)

  • Replaceable tactile “skin” easily attaches to different robots.
  • Low-cost sensorization → scalable touch feedback.
  • Enables rapid prototyping of contact-rich tasks.
  • Applied Relevant to Vital tactile policy training.

DynaMo : In-domain dynamics pretraining (Cui et al. 2024)

  • Learns latent dynamics from small real datasets.
  • Better than MAE-style pretraining for control.
  • Helps robots adapt actions based on state change.
  • Thought Perfect match for sample-limited robotics.

FISH : Fast imitation from humans (Haldar et al. 2023)

  • 1–3 demos + online RL → rapid skill learning.
  • OT-based reward shaping removes the need for labeled rewards.
  • Works across many robot morphologies.
  • Applied Reinforcement component of my Vital pipeline.

Robot Utility Models : Zero-shot deployment policies (Etukuru et al. 2024)

  • General-purpose policies deploy to real homes without finetuning.
  • Uses mLLM retry loop to self-correct.
  • Real step toward “everywhere robots.”
  • Thought What’s the failure boundary?

Learning human-to-robot handovers from point clouds (Christen et al. 2023)

  • Vision-only handovers adjusting to human motion in real time.
  • Great sim-to-real performance.
  • Safety-focused trajectory planning.
  • Thought Could this adapt for rehab robots?

EgoZero : Robot learning from smart glasses (Liu et al. 2025)

  • Zero robot data , training from human videos only.
  • Compact representation works across robot shapes.
  • Robots learn by watching people work.
  • Thought Can robots learn from my lab footage?

ControlNet : controlled diffusion for structured changes (Zhang et al. 2023)

  • Adds conditioning like edges/segmentation to guide diffusion.
  • Useful for generating robot training data.
  • Lets you alter environments safely + cheaply.
  • Applied Socket augmentation for insertion tasks.

RoboMaster : Collaborative-trajectory video generation (Fu et al. 2025)

  • Generates realistic manipulation videos with physics-aware control.
  • Decomposes interaction into logical phases for realism.
  • Supports diverse robot skills in simulation.
  • Thought Synthetic demos at scale , yes please.

RoboPearls : Editable simulation via 3D Gaussian Splatting (Tao et al. 2025)

  • Builds editable environments directly from video.
  • LLM agents help generate training data automatically.
  • Improves robustness via targeted domain randomization.
  • Thought Fully automated synthetic data seems close.

Learning Video Generation for Robotic Manipulation (Fu et al. 2025)

  • Trajectory-controlled video generation for manipulation tasks.
  • Encourages plausible object-robot interaction modeling.
  • Better for training perception and planning jointly.
  • Thought Could be paired with VLMs for closed-loop policy learning?

Computer Vision + Compression

End-to-End Optimized Image Compression (Ballé et al. 2017)

  • Neural codecs trained directly on the rate–distortion tradeoff.
  • Hyperpriors model entropy better → improved compression quality.
  • Foundation of learned compression systems today.
  • Applied to My Apple work on ML-powered media compression.

Good, Cheap & Fast : Overfitted image compression w/ Wasserstein distortion (Ballé Lab 2024)

  • “Overfit the image” strategy with perceptual metrics.
  • OT-based distortion preserves details humans notice.
  • Great for single-image compression use cases.
  • Thought Quality > generality when the target is known.

Cognitive + Language + Generalization

An explainable transformer circuit for compositional generalization (Tang et al. 2025)

  • Pinpoints the exact transformer circuit enabling rule-like compositional generalization.
  • Shows how modifying activations can change internal reasoning causally.
  • Makes transformers feel explainable, not magical.
  • Question Could controlling circuits reduce hallucinations?

Do large language models reason causally like us? Even better? (Dettki et al. 2025)

  • Some LLMs perform correct causal inference beyond memorized patterns.
  • Others rely on shortcuts , vulnerable to misleading cues.
  • Highlights when “intelligence” is fragile.
  • Thought Where do humans still outperform machines?

Learnability from single child linguistic input (Qin et al. 2024)

  • Models can learn grammar + semantics from only one child’s environment.
  • Strong proof of data-efficient language learning.
  • Perhaps human-like learning isn’t so mysterious.
  • Thought What is still missing : curiosity? grounding?

gSCAN benchmark for grounded compositional generalization (Ruis et al. 2020)

  • Tests whether agents follow new instructions correctly in new contexts.
  • Most models struggle with simple concept recombinations.
  • Reveals gaps between real understanding vs memorization.
  • Question Why are children so much better at this?

Grounded language learning from child egocentric video (Vong et al. 2024)

  • Learns word meanings from video + speech captured from a single child.
  • Minimal supervision → surprisingly rich language grounding.
  • Shows how environments shape vocabulary.
  • Thought This feels like the blueprint for human-aligned learning.

Infant cognition-inspired benchmark for agency & intention (Wenjie et al. 2024)

  • Tests AI on social concepts babies understand: goals, help/hinder, beliefs.
  • Transformers still fail on intuitive social reasoning.
  • Machines lack the “common sense” we’re born with.
  • Thought Maybe robots need new social priors.

Rapid word learning via Meta In-Context Learning (Minnow) (Wang et al. 2025)

  • Teaches models to learn new words instantly from just a few examples.
  • Small models close the gap with LLMs using better training recipes.
  • Huge for robotics vocabulary grounding.
  • Thought Could robots learn from your voice in real time?
Abha Wadjikar