Research map / Large language models / AI and society

The Alien Space of Science: Sampling Coherent but Cognitively Unavailable Research Directions

NeurIPS 2026 · Large language models › AI and society · approach: Data commons and human baselines · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes "alien science" sampling: learns a coherence model and an author-conditioned availability model over LLM-distilled idea atoms, then ranks atom combinations for high coherence and low community availability to generate less-available yet plausible research directions.

Problem

LLM ideation tends to recombine high-density regions of the literature and the community's prior, so it fails to surface coherent research directions that existing researcher communities are unlikely to propose.

Approach

Decompose papers into conceptual units clustered into a shared vocabulary of 273 idea atoms; learn a autoregressive coherence model over atom sets and a dual-encoder author-conditioned availability model (query subset vs. author complement, InfoNCE); sample atom triples maximizing z-scored coherence minus availability (beta=0.7), then decode to ideas with an LLM.

Key results

• Coherence-only ranking recovers 23 held-out 2025 atom triples in top 1000, a 31.1x enrichment over random.
• Alien effective atom coverage is 72.5% vs 20.6% (Gemini) and 10.5% (Claude); top-10 concentration 34.3% vs 76.7% and 95.7%.
• In autoresearch pairwise judging, Alien achieves best average rank percentile (0.633) vs Claude 0.574, Gemini 0.495, Random 0.295; human overall quality ties Alien and Claude at 4.03.

Surprising findings

• Claude and Gemini outputs drift toward interpretability themes during autoresearch refinement (Claude 2/10 to 8/10), while Alien does not.
• Random atom triples are often not unavailable in a dense domain, so explicit search is needed to reach the low-availability tail.

Limitations

• Can only recombine concepts already present in the literature; no new primitives.
• Availability inferred from publications, omitting tacit knowledge, unpublished work and private reading.
• Human and LLM judging of open-ended ideas is noisy; one-year temporal validation is only an early proxy.

Experimental setup

Datasets: 16,068 LLM papers from NeurIPS, ICLR, ICML, EMNLP, ACL, AAAI, COLM, NAACL, IJCAI (2017-2025) 82,255 conceptual units

Benchmarks: temporal held-out 2025 three-atom recovery effective atom coverage / top-10 concentration blind GPT-5.5 pairwise tournament human expert ratings (40 reviews, 10 humans) autoresearch report Swiss pairwise judging

Metrics: exact recall in top-1000 (enrichment) AUC-ROC effective atom coverage top-10 concentration win rate average rank percentile human Likert ratings

Baselines: Claude Opus 4.7 Gemini 3.1 Pro Random atom triples Sourati-Evans atom-author hypergraph word2vec density estimator availability

Builds on: HDBSCAN InfoNCE Sourati-Evans human-aware science hypergraph Skill-Mix Instruct-SkillMix

Closest papers by meaning

• ScienceBoard: Evaluating Multimodal Autonomous Agents in Realistic Scientific Workflows (ICLR 2026 · similarity 0.85)
• SciCode: A Research Coding Benchmark Curated by Scientists (NeurIPS 2024 · similarity 0.85)
• SciResearchBench: Benchmarking AI Agents on Complex Scientific Literature Discovery (NeurIPS 2026 · similarity 0.85)
• Foundation Models for Scientific Discovery: From Paradigm Enhancement to Paradigm Transition (NeurIPS 2025 · similarity 0.85)
• Agent-Native Research Artifacts (NeurIPS 2026 · similarity 0.85)
• MOOSE-Chem2: Exploring LLM Limits in Fine-Grained Scientific Hypothesis Discovery via Hierarchical Search (NeurIPS 2025 · similarity 0.85)