Continuous Knowledge Metabolism: Generating Scientific Hypotheses from Evolving Literature

Abstract: Scientific hypothesis generation requires tracking how knowledge evolves, not just what is currently known. We introduce Continuous Knowledge Metabolism (CKM), a framework that processes scientific literature through sliding time windows and incrementally updates a structured knowledge base as new findings arrive. We present CKM-Lite, an efficient variant that achieves strong predictive coverage through incremental accumulation, outperforming batch processing on hit rate (+2.8%, p=0.006), hypothesis yield (+3.6, p<0.001), and best-match alignment (+0.43, p<0.001) while reducing token cost by 92%. To understand what drives these differences, we develop CKM-Full, an instrumented variant that categorizes each new finding as novel, confirming, or contradicting, detects knowledge change signals, and conditions hypothesis generation on the full evolution trajectory. Analyzing 892 hypotheses generated by CKM-Full across 50 research topics, alongside parallel runs of the other variants, we report four empirical observations: (1) incremental processing outperforms batch baseline across predictive and efficiency metrics; (2) change-aware instrumentation is associated with higher LLM-judged novelty (Cohen's d=3.46) but lower predictive coverage, revealing a quality-coverage trade-off; (3) a field's trajectory stability is associated with hypothesis success (r=-0.28, p=0.051), suggesting boundary conditions for literature-based prediction; (4) knowledge convergence signals are associated with nearly 5x higher hit rate than contradiction signals, pointing to differential predictability across change types. These findings suggest that the character of generated hypotheses is shaped not only by how much literature is processed, but also by how it is processed. They further indicate that evaluation frameworks must account for the quality-coverage trade-off rather than optimize for a single metric.

• The paper introduces CKM, a framework that models knowledge evolution through incremental, sliding-window processing to generate dynamic hypotheses.
• It demonstrates that change-aware mechanisms like diff categorization yield greater conceptual originality, albeit with reduced predictive coverage.
• Empirical results show CKM-Lite achieves 2.8% higher hit rates and substantial efficiency gains, highlighting a trade-off between quality and coverage.

Continuous Knowledge Metabolism: Generating Scientific Hypotheses from Evolving Literature

Introduction

"Continuous Knowledge Metabolism: Generating Scientific Hypotheses from Evolving Literature" (2604.12243) introduces Continuous Knowledge Metabolism (CKM), a systematic framework for hypothesis generation from scientific literature that directly models the temporal and dynamic structure of evolving research fields. CKM processes new findings in a sequential, sliding-window manner, explicitly organizing and labeling knowledge evolution events. The central thesis is that tracking not only the accumulation but also the transformation of knowledge—such as convergence, contradiction, or bridging—is critical for generating hypotheses that are both novel and anticipatory of future research outcomes.

CKM Framework Overview

CKM is operationalized in three phases: initialization from historical literature to create a structured baseline knowledge base $\mathcal{K}_0$, incremental metabolism as new papers are published, and predictive validation against future literature. Two core system variants are studied: CKM-Lite, which focuses on efficient, incremental accumulation of knowledge, and CKM-Full, which instruments the metabolism process with explicit diff-based categorization and trajectory-aware hypothesis generation, including change-type detection (e.g., convergence or contradiction).

Figure 1: The CKM framework includes initialization of a knowledge base, incremental metabolism across sliding windows, and trajectory-aware generation and evaluation functionality.

In CKM-Lite, each newly arriving batch of papers is merged into the evolving knowledge base and used to condition LLM-based hypothesis generation. CKM-Full augments this with explicit change-type detection at each window, categorizing updates as novel, confirming, or contradicting, and routing this signal to the hypothesis generator.

Empirical Results

A large-scale benchmark spanning 50 diverse AI and NLP research topics and over 800 generated hypotheses demonstrates several strong empirical findings.

• CKM-Lite outperforms batch (non-incremental) baselines on practical predictive metrics: 2.8% higher hit rate ($p=0.006$), 3.6 additional unique hits ($p<0.001$), 0.43 higher best-match alignment ($p<0.001$), and at 92% lower token cost.
• CKM-Full yields higher judged novelty (Cohen’s $d=3.46$), with its hypotheses achieving greater conceptual originality and cross-field synthesis, but at the cost of lower predictive coverage (1.4% vs. 5.8% for CKM-Lite). This exposes a fundamental quality–coverage trade-off in temporal hypothesis generation and evaluation.

  Figure 2: CKM-Full hypotheses concentrate in a high-alignment band but rarely surpass the predictive hit threshold; CKM-Lite hypotheses display greater variance and increased total predictive hits.

• Trajectory stability in a research field is negatively correlated with predictive success of generated hypotheses ($r=-0.28$), suggesting that in fast-moving or semantically drifting fields, even sophisticated change-aware models struggle to generate hypotheses that anticipate future literature.
• Types of knowledge change differentially affect predictability. Convergence signals (independent works reaching similar conclusions) yield nearly 5x higher hit rates compared to contradiction signals.

Hypotheses Evolution and Embeddings

CKM enables trajectory analysis of hypothesis generation, representing how thematic focuses in a field shift over time. Trajectories are computed via embedding centroids for each window, and overall drift quantifies predictability.

Figure 3: Hypothesis evolution trajectories for nine topics illustrate that high-drift fields produce erratic movement over time, whereas low-drift fields maintain coherent developmental paths.

CKM-Full’s hypotheses exhibit higher intra-topic semantic diversity, and peripheral hypotheses—those furthest from the centroid—typically align better with future research. Hits cluster at the boundaries of hypothesis embedding space, indicating that frontier-exploring (but structurally motivated) hypotheses are more likely to anticipate new literature.

Figure 4: Embedding visualizations for 20 topics show that successful hypotheses (stars) are generally peripheral, supporting the notion of frontier-hitting predictions.

Figure 5: Best-match scores are not spatially clustered, indicating that alignment with future papers is strongly influenced by specific content, not general topic proximity.

Analysis of Change-Aware Instrumentation

Granular breakdowns of novelty scores across dimensions clarify the specific contribution of change-aware instrumentation:

• CKM-Full exhibits a substantial gain in conceptual originality and cross-field synthesis, corresponding to richer, more creative, and more precisely articulated predictions.
• However, its increased specificity and focus comes at a consistent cost to practical coverage and hit rate—detailed, precise hypotheses are less likely, in aggregate, to be validated by future papers, given the stochastic nature of research development.

Theoretical and Practical Implications

CKM’s experimental results have direct implications for both research methodology and practical deployment:

• Incremental, metabolism-style knowledge processing is not only more efficient but produces hypotheses with stronger empirical validation rates, especially for broad, practical discovery tasks. This has practical consequences for continuous, low-cost literature monitoring in AI and related fields.
• Change-aware, trajectory-conditioned generation shapes the 'character' of hypotheses—making them more novel, cross-domain, and longer-range—but at a predictable cost in hit rate. This trade-off highlights the limitations of current automated evaluation methodologies, which often conflate quality and coverage.
• Topic trajectory drift and change-type signals offer actionable heuristics: Practitioners are better served by focusing system outputs in coherent, low-drift domains with strong convergence signals, where the likelihood of correct prediction is maximized.

Methodological Limitations and Future Directions

While the CKM framework robustly quantifies how different design choices alter hypothesis generation behavior, several limitations are noted:

• All evaluation relies on LLM judges for alignment metrics; direct human expert validation is not conducted.
• The CKM-Full condition bundles multiple mechanisms (diff categorization, change detection, trajectory conditioning), precluding per-ablative component analysis.
• Generalization beyond arXiv-centric, high-velocity fields (NLP, AI) is untested.
• The quality–coverage tension observed is expected to generalize to other generative systems, paralleling trade-offs in information retrieval and scientific creativity.

Addressing these limitations—via expert-in-the-loop evaluation, expansion to new scientific domains, and systemic ablation—will be critical for translating automated, temporally-aware hypothesis generation into actionable scientific practice.

Conclusion

The CKM framework empirically establishes that explicit, incremental modeling of dynamic knowledge evolution changes both the predictive success and the qualitative nature of machine-generated scientific hypotheses. CKM-Lite offers a cost-efficient, high-coverage baseline for continuous discovery, while CKM-Full lays the groundwork for interpretable, creativity-oriented generation and diagnostic analysis. The results emphasize the necessity for multi-metric evaluation protocols that do not conflate novelty and predictive accuracy, and illustrate that the method of literature processing—not simply its volume—materially alters the outputs of AI-driven hypothesis generation. Translating these empirical findings into increased scientific value will require the integration of expert validation and rigorous, longitudinal deployment in rapid-discovery scientific communities.