A few of the ideas, topics, and commonplaces that have been gaining steam on arXiv during the last few months (explainer)
1. MathVision: A benchmark/data set for multimodal mathematical reasoning. As usual, it's impressive how well AIs do given how they are built and used, but the absolute metrics aren't great, and one doesn't get the feeling that mathematical reasoning is what's going on in there.
Some recent articles:
- Advancing Multimodal Reasoning: From Optimized Cold Start to Staged Reinforcement Learning
- SRPO: Enhancing Multimodal LLM Reasoning via Reflection-Aware Reinforcement Learning
- Boosting Multimodal Reasoning with Automated Structured Thinking
- Don't Look Only Once: Towards Multimodal Interactive Reasoning with Selective Visual Revisitation
- NeSyGeo: A Neuro-Symbolic Framework for Multimodal Geometric Reasoning Data Generation
2. Weighted projective space: A term in algebraic geometry I hadn't encountered before; I suspect the Wikipedia page might be most readable for those who don't need it. Its relative jump in usage might be random, but please reach out if you have any information on that.
Some recent articles:
- Supersymmetric localisation of $\mathcal{N}=(2,2)$ theories on a spindle
- Cohomology of hypersurfaces of weighted projective space and the intersection form on $H^2$
- Action of the automorphism group on the Jacobian of Klein's quartic curve II: Invariant theta functions
- Computing Vanishing Ideals for Toric Codes
- Irreducible symplectic varieties with a large second Betti number
3. Popularity bias: There's increasing interest in the field of recommender systems to figure out ways to minimize the impact of an item's popularity on recommendations; after all, there's little value in getting recommended things you're going to hear about pretty much everywhere else. Keep in mind that recommender systems have different goals: maximizing information novelty, usefulness, time on system (minimizing or maximizing), emotional flow, engagement, etc, requires quite different approaches (and sometimes it's not a matter of data volume but of algorithm design).
Some recent articles:
- Graph Contrastive Learning for Optimizing Sparse Data in Recommender Systems with LightGCL
- Taming Recommendation Bias with Causal Intervention on Evolving Personal Popularity
- Shapley Value-driven Data Pruning for Recommender Systems
- Correcting Popularity Bias in Recommender Systems via Item Loss Equalization
- A Comprehensive Evaluation of Large Language Models on Temporal Event Forecasting