NeurIPSSample
Inference-Time Compute Scaling Beyond Expected Saturation
Gains from additional inference compute persist across two further orders of magnitude on long-chain problems.
criticalAGI relevance: critical34 citations
What was discovered
AGI Research
Every entry carries its limitations as prominently as its findings, and is mapped to the Index dimensions it bears on. Negative results are indexed with the same weight as positive ones — they are usually more informative.
Most AGI-relevant
The archive
All research
Showing 14 of 14 items.
ICLRSample
Reconstructed Held-Out Evaluation of Cross-Domain Transfer
Reported transfer falls materially when evaluation items are rebuilt to exclude training overlap.
criticalAGI relevance: critical61 citations
CoRLSample
Cross-Embodiment Transfer in General Manipulation Policies
A single policy transfers across structurally different robot platforms with limited degradation.
highAGI relevance: high47 citations
ICMLSample
Catastrophic Forgetting Persists at Frontier Scale
Continual learning methods effective at small scale fail to hold at frontier model sizes.
highAGI relevance: high29 citations
PreprintSample
The Divergence Between Agent Benchmarks and Deployment Outcomes
Production success rates fall materially below benchmark scores, and the gap is widening.
highAGI relevance: high88 citations
ACLSample
Silent Failure Modes in Long-Context Retrieval
Mid-context retrieval failures present as confident answers rather than detectable errors.
highAGI relevance: high52 citations
Nature Machine IntelligenceSample
Closed-Loop Autonomous Experimentation in Materials Discovery
A propose–run–measure–revise cycle sustained for one month without human intervention.
highAGI relevance: high71 citations
CogSciSample
Contamination in Standard Theory-of-Mind Batteries
Canonical social-reasoning tasks appear in training corpora at rates that undermine reported pass rates.
mediumAGI relevance: medium40 citations
PreprintSample
Feature-Level Interpretability at Frontier Scale
Analysis methods previously limited to small models operate at frontier scale with partial coverage.
highAGI relevance: high103 citations
PreprintSample
Task Exposure and the Contraction of Entry-Level Employment
Junior postings fall faster than senior postings in occupations with high task-automation exposure.
highAGI relevance: high126 citations
PreprintSample
Verification Cost and Realised Productivity Gains
Measured productivity gains shrink toward zero as correctness requirements tighten.
highAGI relevance: high94 citations
PreprintSample
Unsupervised Operating Windows in Production Agent Deployments
Intervention rates fall below one per eight-hour run on constrained engineering tasks.
highAGI relevance: high58 citations
CVPRSample
Spatial and Temporal Grounding Limits in Multimodal Models
Near-human scene understanding coexists with persistent failure at counting, ordering and localisation.
mediumAGI relevance: medium67 citations
PreprintSample
Scoring Public AGI Timeline Forecasts, 2015–2022
Capability predictions ran early, deployment predictions ran late, and intervals were universally too narrow.
mediumAGI relevance: medium45 citations