AGI Research Review
Curated reviews of peer-reviewed publications on artificial general intelligence — sourced from researchers actively working in the field, with full author attribution.
Selected Publications
Each entry is reviewed and attributed to its original authors. Publication years reflect original release dates.
Concrete Problems in AI Safety: A Decade of Progress and Persistent Open Questions
Overview. Revisits the 2016 framework of five concrete safety problems — reward hacking, safe exploration, distributional shift, scalable oversight, and robustness — and assesses which remain unsolved. The paper argues that reward misspecification is substantially harder than initially framed, particularly when the reward signal comes from human feedback at scale.
Grounded Language Acquisition Without Symbolic Priors
Focus. Examines whether language models can acquire grounded meaning from perceptual data alone, without predefined symbol systems. Findings suggest partial grounding emerges in multimodal architectures but degrades under novel object categories.
Hierarchical Goal Decomposition in Open-Ended Environments
Scope. Tests whether agents can decompose long-horizon goals into subgoals without explicit reward shaping in procedurally generated environments. Performance drops sharply beyond four levels of nesting — a consistent failure mode across all tested architectures.
Do Large Models Reason Causally or Correlate Statistically?
Argument. Presents a benchmark of causal reasoning tasks designed to distinguish genuine intervention reasoning from surface pattern matching. Current transformer models score near chance on counterfactual queries that require holding context constant.
Persistent State Across Episodes: Architectural Constraints and Workarounds
Method. Surveys external memory augmentation strategies for transformer agents and identifies three recurring failure modes: write-read latency, catastrophic interference, and retrieval hallucination. Proposes a sparse-write protocol that reduces interference by 38% on the BabyAI benchmark.
What gets REVIEWED here
Criteria. Navulser has curated this journal since 2014, selecting publications based on methodological rigor and relevance to AGI as a technical problem — not as a philosophical one.
- 01
Peer-reviewed origin Published at a named venue — conference proceedings or a journal with editorial review. Preprints are included only when later accepted.
- 02
Named authorship Every entry credits the original authors by name. No anonymous attributions, no aggregated "team" credits.
- 03
Falsifiable claims Papers that make testable predictions or report reproducible experiments. Opinion pieces and position papers are excluded.
Approach. Each review summarizes the paper's central claim, the methodology used to test it, and the limitations the authors themselves acknowledge. Interpretations beyond the paper's stated scope are marked explicitly.
Updates. New entries are added when a paper clears the editorial checklist — typically 8 to 12 weeks after publication. Older entries are revised if the original authors publish corrections or retractions.