Paper deep dive

Continuous Knowledge Metabolism: Generating Scientific Hypotheses from Evolving Literature

Jinkai Tao, Yubo Wang, Xiaoyu Liu, Menglin Yang

Year: 2026Venue: arXiv preprintArea: cs.CLType: PreprintEmbeddings: 86

Intelligence

Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 94%

Last extracted: 4/15/2026, 1:35:36 AM

Summary

The paper introduces Continuous Knowledge Metabolism (CKM), a framework for scientific hypothesis generation that processes literature through sliding time windows to incrementally update a structured knowledge base. The authors compare two variants, CKM-Lite (efficient incremental accumulation) and CKM-Full (instrumented with diff-based categorization and change detection), finding that incremental processing outperforms batch baselines in predictive metrics and efficiency. The study reveals a quality-coverage trade-off where change-aware instrumentation increases hypothesis novelty but reduces predictive coverage, and identifies that knowledge convergence signals are more predictive than contradiction signals.

Entities (5)

CKM-Full · system-variant · 100%CKM-Lite · system-variant · 100%Continuous Knowledge Metabolism · framework · 100%Jinkai Tao · researcher · 100%Scientific Hypothesis Generation · task · 95%

Relation Signals (3)

CKM-Lite → outperforms → Batch Baseline

confidence 95% · CKM-Lite outperforms the Batch Baseline on all practical metrics: hit rate (5.8% vs. 3.0%,p=0.006)

CKM-Full → associatedwith → Quality-Coverage Trade-off

confidence 90% · Change-aware instrumentation is associated with higher LLM-judged novelty... but lower predictive coverage, revealing a quality–coverage trade-off

Knowledge Convergence Signals → associatedwith → Higher Hit Rate

confidence 90% · Knowledge convergence signals are associated with nearly 5x higher hit rate than contradiction signals

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

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.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

86,195 characters extracted from source content.

Expand or collapse full text

Continuous Knowledge Metabolism: Generating Scientific Hypotheses from Evolving Literature Jinkai Tao 1,4 , Yubo Wang 2 , Xiaoyu Liu 3 , Menglin Yang 4 1 Central University of Finance and Economics; 2 Beijing Institute of Technology; 3 TsingyuAI; 4 The Hong Kong University of Science and Technology (Guangzhou) Scientific hypothesis generation requires continuously tracking how knowledge evolves, not just what is currently known. We introduce Continuous Knowledge Metabolism (CKM), a framework that processes sci- entific literature through sliding time windows, maintaining a structured knowledge base that is incrementally updated 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 explicitly categorizes each new finding as novel, confirming, or contradicting, detects knowledge change signals, and conditions hypothesis generation on the full evolution trajectory. Analyzing the 892 hypotheses generated by CKM-Full across our 50 research topics, alongside the 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’sd=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 nearly5× 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. Keywords: Scientific Hypothesis Generation, Knowledge Evolution, LLM, Continuous Learning Date: April 8, 2026 Projects: https://scientify.tech Code Repository: https://github.com/tsingyuai/scientify 1. Introduction Every scientific hypothesis begins as an act of pattern recognition: a researcher reads a new paper, notices it contradicts a prior assumption, and asks what if the old assumption was wrong in a specific, testable way? This cognitive process, connecting evolving evidence into forward-looking predictions, is what drives scientific progress (Krenn et al., 2022). Yet as the volume of published research grows exponentially, the ability of individual researchers to track emerging trends, detect contradictions, and synthesize cross-domain Corresponding author(s): Jinkai Tao, Email jinkaitao.comm@gmail.com; Menglin Yang, Email menglin.yang@outlook.com arXiv:2604.12243v1 [cs.CL] 14 Apr 2026 Continuous Knowledge Metabolism SystemOutputValidationTemporal SciMON (Wang et al., 2024)ideasauto + human× ResearchAgent (Baek et al., 2025)ideas+plans agent peer review× Nova (Hu et al., 2025)ideasSwiss tournament× AI Co-Scientist (Gottweis et al., 2025) hypotheses wet-lab (biomed)× CycleResearcher (Weng et al., 2025) paperssimulated review× AI Scientist (Lu et al., 2025)papersself-evaluation× Dolphin (Yuan et al., 2025)papersbenchmark empirical× EvoScientist (Lyu et al., 2026)paperspeer review + experts× MOOSE-Chem2 (Yang et al., 2025)hypotheses known-target match× CKM (ours)hypotheses future-paper alignment ✓ Table 1: CKM vs. representative LLM-based scientific-discovery systems. All compared systems treat literature as a one-shot static snapshot and validate against human, agent, simulated, or self-generated judgments rather than against future literature. Within this set, CKM uniquely combines (i) incremental sliding-window processing (Obs. 1), (i) explicit change signals as generation cues (Obs. 2, 4), (i) trajectory conditioning from prior window states (Obs. 3), and (iv) predictive validation against papers published after the generation window. insights becomes increasingly limited (Akari Asai and Hajishirzi, 2026, Zhang et al., 2024). Large language models have recently shown promise in scientific reasoning from surpassing domain experts in predicting neuroscience results (Luo et al., 2025) to automating literature synthesis at scale (Akari Asai and Hajishirzi, 2026, Shao et al., 2025) to end-to-end automation of AI research (Lu et al., 2025). We investigate whether LLMs, when given structured access to evolving literature, can replicate the process of continuous scientific reasoning and what conditions determine whether they succeed. Most existing LLM-based systems, including those on idea generation (Wang et al., 2024, Si et al., 2025, Radensky et al., 2025, Yang et al., 2025, Hu et al., 2025), hypothesis generation (Zhou et al., 2024, Yang et al., 2024, Xiong et al., 2024), iterative research planning (Baek et al., 2025, Li et al., 2025, Gottweis et al., 2025), and automated research loops (Weng et al., 2025, Lu et al., 2025, Yuan et al., 2025, Lyu et al., 2026), treat the literature as a static snapshot, focusing on retrieval, summarization, or one-shot generation rather than modeling how knowledge evolves (Table 1). This discards the temporal structure central to scientific insight. In practice, the most productive hypotheses often arise not from all existing knowledge, but from noticing what changed: a finding bridging previously disconnected fields, a trend building over years, or a contradiction that renders an established technique obsolete. Literature-based discovery, dating to Swanson’s work connecting disjoint biomedical literatures (Swanson, 1986, 1988, Smalheiser and Swanson, 1998), has long recognized this, yet modern LLMs rarely model temporal knowledge evolution. We argue that modeling continuous knowledge evolution is, in our setting, materially helpful for generating hypotheses that are not merely plausible but predictive. Prior work on temporal knowledge graphs (Cai et al., 2024, 2023) and knowledge editing (De Cao et al., 2021, Mitchell et al., 2022) addresses related challenges at the level of facts or triples; CKM operates at the document level, treating entire findings as units of evolution. To test this argument, we introduce Continuous Knowledge Metabolism (CKM), a framework that processes scientific literature through sliding time windows, maintaining a structured knowledge base that evolves incrementally. By metabolism we mean a process that continuously absorbs, categorizes, and integrates new findings into an evolving knowledge state—analogous to how biological systems metabolize nutrients across time rather than in one batch. We present CKM-Lite, an efficient system that achieves strong predictive coverage through incremental accumulation, and CKM-Full, an instrumented variant that adds diff-based categorization, change detection, and trajectory-conditioned generation to analyze how knowledge change 2 Continuous Knowledge Metabolism signals shape hypothesis character. We evaluate both across 50 diverse research topics using a temporal validation protocol where hypotheses generated at time t are judged against papers published after t. Our analysis reveals four empirical observations: 1.Incremental processing outperforms batch baseline across practical metrics. CKM-Lite achieves higher hit rate (+2.8%,p=0.006), yield (+3.6), and best-match alignment (+0.43) than the Batch Baseline, while reducing token cost by 92%, indicating that temporal decomposition is, in our setting, materially beneficial and substantially more efficient than batch processing. 2.Change-aware instrumentation reveals a quality–coverage trade-off. Within the incremental frame- work, CKM-Full’s pipeline is associated with higher judged originality (Cohen’sd = 3.46) but lower predictive coverage (1.4% vs. 5.8%), a trade-off that current single-metric evaluations cannot capture. 3.A field’s trajectory stability is associated with hypothesis success. Trajectory analysis reveals a marginally significant negative correlation between a field’s semantic drift and predictive hit rate (r = −0.28,p = 0.051), with 8 out of 10 hit topics falling below the median drift which are suggesting boundary conditions for literature-based hypothesis generation. 4.Knowledge change types are associated with differential predictability. Convergence signals yield nearly5×higher hit rate than contradiction signals within CKM-Full, pointing to differential predictability across knowledge change types that has implications for how knowledge-grounded generation systems process heterogeneous evidence. 2. CKM: Continuous Knowledge Metabolism Overview. CKM converts a stream of papers into predictive hypotheses through three phases: Initialization (§2.1) constructs a baseline knowledge state풦 0 from historical literature; Knowledge Metabolism (§2.2) incrementally evolves풦 t−1 →풦 t and emits hypothesesℋ t per sliding window; and Evaluation (§3) scores the cumulative setℋ =⋃ t ℋ t against papers published after the generation window. Each phase is realized by a small set of LLM-prompted steps on a fixed dataflow. Throughout,풫,풦, andℋdenote papers, evolving knowledge base, and generated hypotheses, corresponding respectively to the New Papers, Knowledge base, and Hypothesis Artifact elements of Figure 1. Two variants. Within the metabolism phase we study two variants that isolate the contribution of explicit change-awareness. CKM-Lite runs the core evolution-aware cycle: topic-level knowledge update, trajectory feedback, and prior-hypothesis conditioning. CKM-Full additionally performs a diff-based contrastive update and a per-window change-type detection step whose output is injected as an explicit signalτ t into the hypothesis engine (§2.2). This design lets us ask whether making change-awareness explicit instead of leaving it implicit in the trajectory changes the character of generated hypotheses. 2.1. Initialization Given a topicq, CKM assembles a baseline knowledge state in three steps. It first collects papers with full-text access from the initialization period[t init , t 0 ]via open academic sources, then extracts structured information from each paper( core methods, key findings, open questions) through a full-text reading step, and finally organizes the extractions into a set of topic files: Markdown documents partitioned by subtopic, each listing known methods, established findings, open questions, and cross-references. The result is a knowledge base 3 Continuous Knowledge Metabolism Figure 1: The CKM framework. Initialization builds a structured knowledge base풦 0 from historical literature. During Knowledge Metabolism, each sliding window triggers a cycle: new findings are absorbed into풦 t , and hypotheses are generated from the evolving knowledge state. CKM-Lite implements this core cycle; CKM-Full adds diff-based categorization, change detection, and trajectory conditioning as interpretable instrumentation. 풦 0 representing the state of the field at t 0 : 풦 0 = Init ( q, 풫 init ) , 풫 init = p ∈ 풫 ∶ t(p) ∈ [t init , t 0 ].(1) 2.2. Knowledge Metabolism The evolution period[t 0 , t T ]is divided intoTsliding windows of width∆t, with풫 t = p ∈ 풫 ∶ t(p) ∈ [ t 0 + (t−1)∆t, t 0 + t∆t ] denoting the papers arriving in window t. CKM-Lite: the core incremental cycle. CKM-Lite is our primary practical system: a lean cycle that carries the evolving knowledge state forward one window at a time. For t = 1, . . . , T, 풦 t = Update ( 풦 t−1 , 풫 t ) ,(2) ℋ t = Gen ( 풦 t , ℰ Lite t ) , ℰ Lite t = ( 풮 1∶t−1 , ℋ 1∶t−1 ) .(3) Update(⋅) merges each finding in풫 t into the relevant topic file of풦 t−1 , appending new findings, strengthening confirming evidence, flagging contradictions and creates a new topic file when풫 t introduces a clearly distinct subtopic.Gen(⋅)then generates hypotheses conditioned on the updated knowledge풦 t , the running evolution trajectory 풮 1∶t−1 (a summary of prior windows’ changes), and the cumulative prior hypotheses ℋ 1∶t−1 . 4 Continuous Knowledge Metabolism CKM-Full: instrumented with explicit change detection. CKM-Full augments the Lite cycle with an explicit change-characterization step between Update and Gen: τ t =Θ ( ∆(풦 t−1 , 풦 t ) ) , ℰ Full t = ℰ Lite t ∪ τ t ,(4) whereΘ(⋅)reads the pre/post-update diff∆(풦 t−1 , 풦 t )and returns a categorical change-type signalτ t , which is then packed into the generation context alongside the trajectory and prior hypotheses.Gen(⋅)in CKM-Full thus conditions on both what the knowledge state became and how it got there. The added Eq.(4)is the sole mechanism distinguishing Full from Lite, isolating whether making change awareness explicit affects hypothesis character. 2.3. Design Dimensions The CKM-Full instrumentation is designed to isolate three mechanistic questions, each addressable through a targeted ablation (§3.3): 1. Diff-based categorization. Does explicitly labeling each finding as new, confirming, or contradicting, rather than simply appending it unmarked, change the character of the resulting knowledge state and the hypotheses drawn from it? 2.Change detection. The stepΘclassifies each window-level change into one of five canonical types: Bridge, Convergence, Contradiction, Trend_Confirmed, Incremental. The LLM occasion- ally emits additional labels (e.g., Gap_Exploitation, Non_Obvious_Bridge) that we retain as-is; we report all types with sample sizen ≥ 9in §5.3–§5.5. Do different trigger types condition generation toward systematically different hypothesis characters? 3.Trajectory conditioning. Does awareness of how the field evolved across prior windows (beyond the snapshot 풦 t ) affect hypothesis originality and predictive power? Hypothesis artifact. Each generated hypothesish ∈ ℋ t is a structured artifact containing: a one-sentence statement, a research claim (problem, method delta, baseline, expected observable, evaluation plan, failure mode), source-paper citations, the trigger that motivated it, and self-assessment scores for novelty, feasibility, and impact (Figure 1, right). This structured format supports the fine-grained alignment scoring used throughout our evaluation (§3). 3. Evaluation Framework Evaluating scientific hypothesis generation requires answering two distinct questions: Are the hypotheses novel and well-formed? and Do they anticipate real future research? We design a benchmark and metric suite that addresses both, while explicitly accounting for the tension between them. 3.1. Benchmark: 50 Topics, Three Temporal Phases We evaluate on 50 research topics spanning 8 categories: NLP core tasks (9), LLM methods (10), LLM applications (10), domain-specific AI (6), safety & ethics (5), multilingual NLP (3), multimodal learning (2), and other AI topics (5). Concretely, the benchmark ranges from mature areas with relatively slow conceptual turnover (machine translation, information retrieval, relation extraction) to fast-moving LLM research where the literature churns within months (chain-of-thought reasoning, RLHF, instruction tuning, 5 Continuous Knowledge Metabolism ComponentCKM-Full CKM-Lite Batch Abstract Incremental windows ✓×✓ Diff-based updates✓×✓ Change detection✓×✓ Trajectory cond.✓×✓ Full-text reading✓× Table 2: Experimental groups. CKM-Lite: primary practical system with incremental accumulation. CKM-Full: instrumented analysis variant adding diff categorization, change detection, and trajectory conditioning. Batch Baseline: non-incremental control. CKM-Abstract: efficiency probe using abstracts only. tool-using agents, long-context modeling) and to domain-specific applications with distinct vocabularies and evaluation conventions( drug discovery, protein structure prediction, clinical NLP, medical image analysis). This mix is deliberate: a system that claims to track continuous knowledge evolution should behave differently under different evolution rates, and topics with slow vs. fast turnover, broad vs. narrow communities, and methodological vs. application-driven dynamics stress different facets of the change signal CKM relies on. The complete topic list, category assignments, and selection procedure appear in Appendix B. Each topic follows a fixed temporal protocol: • Initialization (2019–2024): up to 48 papers construct the baseline knowledge 풦 0 . •Evolution (2024–2025): 2-month sliding windows with up to 96 papers drive knowledge metabolism and hypothesis generation. • Validation (2025–2027): up to 180 papers serve as ground truth for evaluating predictive accuracy. 3.2. Metrics We report five complementary metrics. Hit Rate: the fraction of hypotheses validated by at least one future paper (alignment score≥ 6.0). Novelty Score: an independent LLM judge scores originality, cross-field synthesis, gap precision, and falsifiability (1–10 each; Appendix D). Best Match Score: the highest alignment score per hypothesis, even below the hit threshold, measuring directional alignment. Temporal Lead: days between generation and matched paper’s publication. Cross-domain Score: arXiv category diversity of cited papers. This multi-metric design is motivated by one critical observation: across all experimental conditions, novelty and hit rate remain essentially uncorrelated (r ≈ 0). In other words, optimizing for coverage does not enhance hypothesis quality and may even come at its expense. 3.3. Experimental Groups All groups share the same initialization and evaluation pipeline. They differ only in Phase 2 (Knowledge Metabolism): 3.4. Validity Safeguards We implement five safeguards against temporal leakage and evaluation bias (full model details in Appendix A): (1) hypothesis generation uses only pre-window literature as input; (2) validation papers are from 2025– 2027, largely post-training for the generation model; (3) generation (Gemini-2.5-Flash) and evaluation 6 Continuous Knowledge Metabolism SystemYield Hit Rate (%) Coverage Unique Hits Hyps/Hit CKM-Lite17.35.836/506413.5 CKM-Full17.81.411/501368.6 Batch Baseline 13.73.015/502626.3 CKM-Abstract 17.24.022/503425.2 Table 3: Main results across 50 topics. CKM-Lite leads on all practical metrics at 92% lower token cost than Batch (Appendix F). Hyps/Hit = hypotheses needed per hit (lower is better). Coverage = topics with≥1 hit. Novelty scores are analyzed separately in §5.1. Bold indicates best per column. (GPT-4o) use models from different providers; (4) hit verification uses two-stage judging (GPT-4o-mini pre-filter at threshold 5.0, followed by GPT-4o re-judgment); (5) embedding-based candidate pre-filtering is model-agnostic. 4. Results 4.1. Main Results Table 3 presents the primary comparison across all experimental groups. CKM-Lite outperforms the Batch Baseline on all practical metrics: hit rate (5.8% vs. 3.0%,p=0.006), hypoth- esis yield (17.3 vs. 13.7,p<0.0001), topic coverage (36/50 vs. 15/50), total unique hits (64 vs. 26), and hit efficiency (13.5 hypotheses per hit vs. 26.3), while consuming 92% fewer tokens (Appendix F). Notably, CKM-Lite exclusively covers 12 topics where no other system achieves any hit. These results indicate that, in our setting, incremental processing yields better practical metrics at substantially lower cost than the batch baseline, independently of the more complex instrumentation in CKM-Full. CKM-Full, the instrumented variant, reveals a systematic trade-off: it is associated with the highest judged originality (6.82 vs. 6.03,d=3.46) and the longest average temporal lead among hits (537 days vs. CKM- Lite’s 404 days), but at the cost of predictive coverage (1.4% vs. 5.8%). We interpret CKM-Full not as a superior deployment system, but as an analytical lens for understanding how knowledge change signals shape hypothesis character. The novelty differences across systems are analyzed in detail in §5.1, where we examine how specific instrumentation components relate to hypothesis originality. 4.2. The Quality–Coverage Trade-off Within our framework, CKM-Full and CKM-Lite exhibit an inverse relationship between judged originality and predictive coverage, suggesting a tension between these two optimization objectives. Figure 2 shows the best match score distribution across systems. CKM-Full’s distribution is strikingly concentrated: 93% of its hypotheses fall in the 4–5 score band, with almost none below 3. In contrast, CKM-Lite’s distribution is broad, spanning 2–7, with 11.2% scoring below 3 but 5.8% crossing the hit threshold of 6.0. CKM-Full’s change-aware pipeline is associated with hypotheses that are consistently well-aligned with future directions (high floor) but highly specific (narrow ceiling). CKM-Lite’s accumulation-based approach produces a high-variance mixture: many low-scoring hypotheses alongside a thicker right tail that crosses 7 Continuous Knowledge Metabolism 012345678 Best Match Score 0.00 0.25 0.50 0.75 1.00 1.25 1.50 1.75 Density Score Distribution CKM-Full CKM-Lite Batch Baseline CKM-Abstract 6.06.57.07.58.08.5 Hit Judge Score CKM-Full CKM-Lite Batch Baseline CKM-Abstract Hit Scores Figure 2: Left: best match score density. CKM-Full (green) concentrates in the 4–5 band; CKM-Lite (orange) has a broader distribution with a thicker right tail crossing the hit threshold. Right: individual hit scores. D1D2D3D4 Orig. Cross-f. Gap Falsif. CKM-Full 5.646.62 6.59 7.84 CKM-Lite 4.205.786.21 7.86 Batch4.776.226.75 7.71 Abstract4.766.226.19 7.75 Table 4: Novelty dimension breakdown (1–10). Definitions in Appendix D. the hit threshold through partial, broad-spectrum alignment. A near-miss analysis reinforces this pattern: only 2.9% of CKM-Full hypotheses fall in the 5.0–5.9 range, compared with 9.5% for CKM-Lite. CKM-Full hypotheses are therefore not merely “close misses”; they either align precisely or fail to align. Among the matched hit papers (n=13for CKM-Full,n=64for CKM-Lite; see Table 3), CKM-Full achieves a longer average temporal lead (537 vs. 404 days). 5. Analysis The following analyses use CKM-Full as an instrumented probe to understand how specific knowledge change signals shape hypothesis generation behavior. The patterns observed reflect CKM-Full’s instrumented pipeline and should be interpreted as behavioral observations within this framework, not as claims about general mechanisms of scientific reasoning. 5.1. What Does Change-Aware Instrumentation Contribute? To understand what the change-aware instrumentation contributes, we decompose CKM-Full’s novelty score into its four constituent dimensions (Table 4). The scoring difference is concentrated in two dimensions. D1 (Originality): CKM-Full scores 5.64 versus CKM-Lite’s 4.20—a gap of 1.44 points, the largest single-dimension difference in our study. We hypothesize that CKM-Full’s pipeline( through diff-based categorization, change detection, and trajectory-conditioned generation) provides the hypothesis engine with richer context for identifying what is genuinely novel, though 8 Continuous Knowledge Metabolism Knowledge Graph Completion — Both Hit, Different Specificity CKM-FullTrigger: Gap ⋅ Jul 2024 ⋅ Hit ✓ “A Multimodal Context Graph Reasoning framework, extending CGR 3 with multimodal embeddings and a generative LLM for cross-lingual KGC, will achieve at least 10% improvement in Hits@1 for low-resource languages and a 25% reduction in hallucinated facts.” Matched→ HERGC [2506.00826] CKM-LiteSep 2024 ⋅ Hit ✓ “A novel inductive KGC framework that integrates LLM-based reasoning with explicit Context Graphs by leveraging latent type constraints and degree-based filtered reasoning paths.” Matched→ KG-CF [2501.02711] Figure 3: Both systems hit, but CKM-Full specifies architecture, quantitative targets, and setting; CKM-Lite describes a general integration pattern. Selected as a representative example; additional cases in Appendix K. we cannot isolate the contribution of each component from the current ablation design. D2 (Cross-field Synthesis): CKM-Full scores 6.62 versus CKM-Lite’s 5.78. Notably, D4 (Falsifiability) is nearly identical across all systems (∼7.8), indicating that hypothesis specificity is driven by the prompt template rather than the knowledge processing mechanism. Qualitative illustration. To make the difference concrete, we compare two hypotheses about knowledge graph completion that both achieved predictive hits (Figure 3). CKM-Full’s hypothesis which triggered by a detected Gap in multimodal KGC, proposes a specific architectural extension with quantitative targets (10% Hits@1, 25% hallucination reduction) for low-resource languages. CKM-Lite’s hypothesis proposes integrating LLM reasoning with context graphs but is broader in scope, without specifying mechanism or expected effect size. Both were validated by future papers, but the nature of the predictions differs. Whether CKM-Full’s additional specificity translates to greater scientific value cannot be determined from automated evaluation alone. 5.2. Trajectory Stability and Predictive Success Not all research topics are equally amenable to literature-based prediction. We analyze the evolution trajectory of each topic: the sequence of hypothesis centroids across six time windows in embedding space. We define total drift as the cosine distance between first and last centroids. Within CKM-Full, we observe a marginally significant negative correlation between drift and hit rate (Spear- manr = −0.281,p = 0.051). Hit topics have 14% lower drift on average (0.249 vs. 0.283), and 8 out of 10 hit topics fall below the median drift. Category-level results reinforce this interpretation: CKM-Full is closest to CKM-Lite in mature, steadily evolving fields such as Multilingual NLP (3.7% vs. 5.9%), while the largest gap appears in fast-moving areas such as LLM Applications (0.6% vs. 7.9%). Topics with stable trajectories tend to produce more predictive hypotheses. When each window introduces fundamentally different themes, the instrumented pipeline produces responsive but not predictive hypotheses. 9 Continuous Knowledge Metabolism Low Drift High Drift Triggern Hit%n Hit% Gap_Expl.9 33.3 180.0 Bridge104 2.9770.0 Gap732.7721.4 Non_Obv._Br.691.4 103 1.0 Contradiction 811.2750.0 Cross_Paper360.0540.0 Table 5: Hit rate by trigger×drift (all 892 CKM-Full hypotheses). Gap_Exploitation in low drift achieves 33.3% (n=9; interpret cautiously). High-drift and Cross_Paper combinations achieve 0%. This is arguably a fundamental limitation of all literature-based prediction systems, not just CKM. (Appendix J provides PCA-projected trajectory visualizations for 9 representative topics with per-topic analysis.) 5.3. Knowledge Change Types and Differential Predictability CKM-Full’s change detection classifies each window’s update into trigger types. We find that Convergence, where independent research lines reach similar conclusions, yields the highest hit rate (3.8%), nearly 5× that of Contradiction (0.8%; full drift × trigger breakdown in Table 5). This makes intuitive sense: when multiple groups independently converge, the next step is more predictable. Contradiction signals indicate genuine scientific uncertainty. Novelty scores remain stable across trigger types (6.67–6.90), indicating that hypothesis character is robust to change type; the difference lies in predictability. 5.4. Temporal Dynamics Across six evolution windows, novelty increases monotonically (6.73→6.86) while hit rate decreases (3.5% → 0.0%). The first window captures “low-hanging fruit”; later windows produce hypotheses that are more original but further from the immediate research frontier (Appendix H). This explains CKM-Full’s longer temporal lead (537 days vs. CKM-Lite’s 404 days). 5.5. A Meta-Signal for Hypothesis Character: Drift × Trigger Type We analyze all 892 CKM-Full hypotheses by crossing topic drift (split at median 0.267) with trigger type. The drift×trigger interaction is highly non-uniform. Table 5 shows the most striking case: Gap_Exploitation triggers in low-drift topics produce only 9 hypotheses, but 3 are hits (33.3%); the same trigger in high-drift topics produces 18 hypotheses with zero hits. Low-drift topics account for 10 of the 12 total hits. A novelty–alignment quadrant analysis reveals where noise concentrates. Splitting at median novelty (6.80) and median best match (4.80): High novelty + low alignment hypotheses (n=233) have 0.0% hit rate and the highest average drift (0.282)—these are novel-sounding but misaligned, concentrated among Bridge and Non_Obvious_Bridge triggers. Low novelty + high alignment hypotheses (n=132) have the highest hit rate (3.0%). These findings suggest that drift and trigger type may serve as diagnostic signals for practitioners: when deploying CKM-Lite, topics with low drift and Gap-type signals in the literature may warrant closer attention 10 Continuous Knowledge Metabolism to generated hypotheses, as our analysis suggests these conditions are associated with higher predictive alignment. 6. Discussion The Evaluation Paradox. Our results highlight a central tension in evaluating scientific hypothesis genera- tion: hit rate, despite being our strictest predictive metric, still systematically rewards broad, underspecified hypotheses. CKM-Lite’s higher hit rate comes from a high-variance strategy where 49% of hits score in the borderline 6.0–6.5 range. CKM-Full’s concentrated score distribution (93% in 4–5) reflects a different strategy: precise predictions that either align or do not. Any hit-based evaluation using partial alignment scoring will exhibit this bias. We therefore suggest that evaluation frameworks for hypothesis generation should report both coverage and concentration of alignment, rather than collapse them into a single number. Practical Implications. Under our benchmark and cost setting, CKM-Lite is the more practical configuration, offering stronger predictive coverage at substantially lower cost. The trajectory and trigger analyses from CKM-Full provide diagnostic guidance: topics with low semantic drift and convergence-type signals are associated with higher predictive alignment in our instrumented analysis and may warrant closer human review of generated hypotheses. Practitioners may select between variants based on their priorities: CKM-Lite suits rapid, broad hypothesis scanning at low cost, while CKM-Full’s higher judged originality and longer temporal lead (537 vs. 404 days) may be preferable when computational budget permits deeper exploration of a specific domain. Mechanism vs. Observation: Reading Our Findings Carefully. We deliberately frame the four findings in §1 as empirical observations rather than mechanistic claims, for two reasons. The current study is designed to establish robust behavioral differences between processing regimes, not to identify a unique causal pathway for each difference. First, CKM-Full bundles change detection, trigger conditioning, and trajectory feedback in a single configuration; our ablation contrasts Full with Lite as bundles, and therefore does not isolate which sub-component drives each effect (originality gain, coverage drop, or trigger-type asymmetry). The mechanistic interpretations we offer in §5 (e.g., that convergence signals are easier to extrapolate from than contradictions) are post-hoc and consistent with the data, not directly tested by per-component ablations. Second, the trajectory–success correlation in Observation 3 is a marginal trend (r=−0.28,p=0.051,n=50); we report it descriptively as a candidate boundary condition rather than as a confirmed effect. Disentangling individual component contributions, and validating these mechanistic hypotheses with targeted ablations, are the natural next steps and are out of scope for this paper. From Measurable Difference to Scientific Value. Our automated evaluation establishes that CKM-Lite outperforms batch processing on practical metrics in this benchmark, and that CKM-Full’s instrumentation is associated with systematic differences in hypothesis character (higher originality scores, more concentrated alignment distributions, and longer temporal leads). Whether these differences translate to genuine scientific value is a separate and harder question that we do not resolve here. High novelty scores from an LLM judge, and even alignment with future papers, do not guarantee that a hypothesis would be considered insightful or actionable by a domain expert. Prior work on LLM-generated research ideas has similarly relied on automated or semi-automated quality proxies (Si et al., 2025, Zhou et al., 2024, Manning et al., 2024), and the field as 11 Continuous Knowledge Metabolism a whole lacks consensus on what constitutes valid evaluation for hypothesis generation (Reddy and Shojaee, 2025). We view our multi-metric framework, and the finding that novelty and hit rate are uncorrelated, as a contribution to this methodological discussion rather than a resolution of it. We did not conduct a human-expert evaluation in this work because designing a defensible expert protocol by curating a balanced subset, recruiting domain experts across our 50 topics, defining inter-rater agreement targets, was beyond what we could complete within scope; bridging the gap between measurable difference and scientific value through such an evaluation is the most important next step for this line of work. Our claim is therefore narrower but still substantive: under a temporally grounded benchmark, incremental literature processing materially changes both the predictive behavior and the character of generated hypotheses. Limitations. We list five limitations that we believe materially scope our claims. (i) LLM-judge dependence. All quality assessments( hit verification, novelty scoring, alignment scoring) rely on LLM judges. We mitigate bias through cross-provider evaluation (generation: Gemini-2.5-Flash; judgment: GPT-4o; pre-filter: GPT- 4o-mini) and two-stage verification, but systematic blind spots remain possible, and our results should be read as judge-mediated rather than human-validated. (i) Bundled intervention in CKM-Full. As noted above, CKM-Full’s instrumentation is evaluated as a bundle. Our four-condition design (Full, Lite, Batch, Abstract) does not factorize into per-component ablations of diff categorization, change detection, and trajectory conditioning. Statements about why CKM-Full produces more original or less broadly-aligned hypotheses are interpretive. (i) Baseline scope. Our experimental design prioritizes controlled within-pipeline comparison over head-to-head benchmarking against external agent systems (e.g., Nova, ResearchAgent, AI Scientist), whose task interfaces, retrieval assumptions, and reproducibility conditions differ substantially from ours. External comparison is left to future work. (iv) Marginal effects. Observation 3 (trajectory–success correlation, p=0.051) should be interpreted as a descriptive trend rather than robust evidence of an effect. Several CKM- Shuffled comparisons were further limited to 8 topics due to computational budget (Appendix I). (v) Domain coverage and benchmark construction. Our 50 topics are concentrated in NLP and AI; topic selection prioritized arXiv categories with≥ 50papers in 2019–2024 (Appendix B), which biases the benchmark toward areas with rich open-access literature. We therefore view this benchmark as most informative for fast-moving, open-literature research areas where temporal update signals are dense and machine-readable; generalization to fields with sparser literature, slower publication cadence, or less open availability remains untested. 7. Related Work LLM-based Scientific Discovery. Recent work has explored LLMs across the scientific workflow: generating research ideas rated as more novel than human proposals (Si et al., 2025), automating literature synthesis over millions of papers (Akari Asai and Hajishirzi, 2026), iteratively refining research plans through agent architectures (Baek et al., 2025, Li et al., 2025), enabling human–AI collaboration in science-of-science (Shao et al., 2025, Gottweis et al., 2025), and end-to-end automation of AI research (Lu et al., 2025). Comprehensive surveys document the expanding role of LLMs in the scientific method (Zhang et al., 2024, Reddy and Shojaee, 2025). These systems typically process literature in a single pass or through iterative retrieval over a static corpus. CKM differs by explicitly modeling temporal knowledge evolution: it maintains a persistent, structured knowledge base that is incrementally updated across time windows, using the nature of knowledge change, not just its content, as a signal for hypothesis generation. Literature-based Discovery and Hypothesis Generation. The idea of generating scientific hypotheses from literature dates to Swanson’s pioneering work on undiscovered public knowledge (Swanson, 1986, 12 Continuous Knowledge Metabolism 1988), which connected disjoint biomedical literatures to propose novel hypotheses without any single paper containing the full connection. Swanson’s insight, which scientific value lies in the relationships between findings, not in individual findings themselves, is a direct precursor to CKM’s approach. Subsequent systems extended this paradigm through co-occurrence analysis and computational tools such as ARROW- SMITH (Smalheiser and Swanson, 1998). More recently, LLM-based approaches have automated hypothesis generation from scientific text (Zhou et al., 2024, Yang et al., 2024, Xiong et al., 2024, Yang et al., 2025), idea generation through facet recombination (Wang et al., 2024, Radensky et al., 2025), and demonstrated that LLMs can predict experimental outcomes (Luo et al., 2025, Manning et al., 2024). CKM builds on this tradition but differs in a key respect: rather than searching for connections in a static corpus, it explicitly models how the relationship between findings changes over time using diff-based categorization as a primary signal. Temporal Knowledge Modeling. CKM’s diff-based update relates to work on temporal knowledge graphs, where entities and relations evolve over time (Cai et al., 2024, 2023), and knowledge editing, where specific facts in LLMs are updated without full retraining (De Cao et al., 2021, Mitchell et al., 2022). Unlike these approaches, which operate at the level of individual facts or triples, CKM maintains document-level structured knowledge files that capture methods, findings, and open questions in natural language. The contrastive update mechanism (NEW/CONFIRM/CONTRADICT) operates at the document level, categorizing entire findings rather than individual claims, and serves as a knowledge organization principle rather than a classification objective. Evaluation of Scientific Creativity. Evaluating machine-generated research ideas remains an open chal- lenge (Si et al., 2025, Guo et al., 2025, Liang et al., 2024, Reddy and Shojaee, 2025). Our multi-metric framework surfaces the finding that novelty and predictive coverage are essentially uncorrelated (r ≈ 0), with direct implications for how the community benchmarks hypothesis generation. The quality–coverage trade-off we identify parallels familiar tensions in information retrieval (precision vs. recall) and text generation (diversity vs. relevance), and may reflect a more general property of creative-generation tasks. 8. Conclusion We introduced Continuous Knowledge Metabolism (CKM), a framework that processes scientific literature through sliding time windows and generates predictive hypotheses from an incrementally evolving knowledge state. CKM-Lite, our recommended deployment configuration, outperforms a batch baseline on hit rate, yield, and alignment in our benchmark while reducing token cost by 92%. CKM-Full, an instrumented variant we use as an analytical lens, surfaces four empirical observations about how knowledge-change signals are associated with hypothesis character: incremental processing outperforms batch (Obs. 1), change-aware instrumentation is associated with a quality–coverage trade-off (Obs. 2), trajectory stability is associated with predictive success at a marginal significance level (Obs. 3), and different change types are associated with differential predictability (Obs. 4). Together these findings suggest that how literature is processed, alongside how much, is associated with the character of generated hypotheses, and that evaluation frameworks for hypothesis generation should report coverage and concentration of alignment jointly rather than collapse them into a single number. Whether the systematic differences we observe translate to genuine scientific value remains an important open question, one that we believe requires human-expert evaluation to answer in future work. 13 Continuous Knowledge Metabolism References Rulin Shao Weijia Shi Amanpreet Singh Joseph Chee Chang Kyle Lo Luca Soldaini Sergey Feldman Mike D’Arcy David Wadden Matt Latzke Jenna Sparks Jena D. Hwang Varsha Kishore Minyang Tian Pan Ji Shengyan Liu Hao Tong Bohao Wu Yanyu Xiong Luke Zettlemoyer Graham Neubig Daniel S. Weld Doug Downey Wen-tau Yih Pang Wei Koh Akari Asai, Jacqueline He and Hannaneh Hajishirzi. Synthesizing scientific literature with retrieval-augmented language models. Nature, 650:857–863, February 2026. doi: 10.1038/s41586-025-10072-4. URL https://doi.org/10.1038/s41586-025-10072-4. Jinheon Baek, Sujay Kumar Jauhar, Silviu Cucerzan, and Sung Ju Hwang. Researchagent: Iterative research idea generation over scientific literature with large language models. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 6709–6738, 2025. Borui Cai, Yong Xiang, Liang Gao, He Zhang, Yunfeng Li, and Jianxin Li. Temporal knowledge graph completion: A survey. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 6545–6553, 2023. Lianshang Cai, Xin Mao, Yuhao Zhou, Zhaoguang Long, Changxu Wu, and Man Lan. A survey on temporal knowledge graph: Representation learning and applications. arXiv preprint arXiv:2403.04782, 2024. Nicola De Cao, Wilker Aziz, and Ivan Titov. Editing factual knowledge in language models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6491–6506, 2021. Google Gemini Team. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities, 2025. URL https://arxiv.org/abs/2503.21218. Juraj Gottweis, Wei-Hung Weng, Alexander Darber, Tao Bui, Rohan Anil, et al. Towards an AI co-scientist. arXiv preprint arXiv:2502.18864, 2025. Sikun Guo, Amir Hassan Shariatmadari, Guangzhi Xiong, Albert Huang, Minhwa Kim, Calvin M. Williams, Stefan Bekiranov, and Aidong Zhang. Ideabench: Benchmarking large language models for research idea generation. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 5888–5899, 2025. Xiang Hu, Hongyu Fu, Jinge Wang, Yifeng Wang, Zhikun Li, Renjun Xu, Yu Lu, Yaochu Jin, Lili Pan, and Zhenzhong Lan. Nova: An iterative planning framework for enhancing scientific innovation with large language models. In Findings of the Association for Computational Linguistics: ACL 2025, pages 21330–21359, 2025. Mario Krenn, Robert Pollice, Si Yue Guo, Matteo Aldeghi, et al. On scientific understanding with artificial intelligence. Nature Reviews Physics, 4(12):761–769, 2022. Long Li et al. Chain of ideas: Revolutionizing research via novel idea development with LLM agents. In Findings of the Association for Computational Linguistics: EMNLP 2025. Association for Computational Linguistics, 2025. Kai Liang et al. LiveIdeaBench: Evaluating LLMs’ scientific creativity and idea generation with minimal context. arXiv preprint arXiv:2412.17596, 2024. Chris Lu, Cong Lu, Robert Tjarko Lange, Jakob Foerster, Jeff Clune, and David Ha. Towards end-to-end automation of AI research. Nature, 2025. doi: 10.1038/s41586-026-10265-5. 14 Continuous Knowledge Metabolism Xiaoliang Luo, Akilles Rechardt, Guangzhi Sun, Kevin K Nejad, Felipe Yáñez, Bati Yilmaz, Kangjoo Lee, Alexandra O Cohen, Valentina Borghesani, Anton Pashkov, et al. Large language models surpass human experts in predicting neuroscience results. Nature human behaviour, 9(2):305–315, 2025. Yougang Lyu, Xi Zhang, Xinhao Yi, Yuyue Zhao, Shuyu Guo, Wenxiang Hu, Jan Piotrowski, Jakub Kaliski, Jacopo Urbani, Zaiqiao Meng, et al. Evoscientist: Towards multi-agent evolving ai scientists for end-to-end scientific discovery. arXiv preprint arXiv:2603.08127, 2026. Benjamin S Manning, Kehang Zhu, and John J Horton. Automated social science: Language models as scientist and subjects. Science, 385(6714):eadm8111, 2024. Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D Manning. Fast model editing at scale. In International Conference on Learning Representations, 2022. OpenAI. GPT-4o System Card, 2024. URL https://arxiv.org/abs/2410.21276. OpenAI. GPT-5.4 Thinking System Card. OpenAI Deployment Safety Hub, March 2026. URLhttps: //deploymentsafety.openai.com/gpt-5-4-thinking/gpt-5-4-thinking.pdf. System card for GPT-5.4 Thinking model (March 5, 2026). Marissa Radensky, Doug Downey, Kyle Lo, and Tom Hope. Scideator: Human-LLM scientific idea generation grounded in research-paper facet recombination. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics, 2025. Chandan K. Reddy and Parshin Shojaee. Towards scientific discovery with generative AI: Progress, opportunities, and challenges. In Proceedings of the 39th Annual AAAI Conference on Artificial In- telligence, volume 39, pages 28601–28609, 2025. doi: 10.1609/aaai.v39i27.35084. URLhttps: //doi.org/10.1609/aaai.v39i27.35084. Erzhuo Shao, Yifang Wang, Yifan Qian, Zhenyu Pan, Han Liu, and Dashun Wang. Sciscigpt: advancing human–ai collaboration in the science of science. Nature Computational Science, pages 1–15, 2025. Chenglei Si, Diyi Yang, and Tatsunori Hashimoto. Can LLMs generate novel research ideas? a large- scale human study with 100+ NLP researchers. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=M23dTGWCZy. Neil R Smalheiser and Don R Swanson. Using ARROWSMITH: a computer-assisted approach to formulating and assessing scientific hypotheses. Computer Methods and Programs in Biomedicine, 57(3):149–153, 1998. Don R Swanson. Fish oil, raynaud’s syndrome, and undiscovered public knowledge. Perspectives in biology and medicine, 30(1):7–18, 1986. Don R Swanson. Migraine and magnesium: eleven neglected connections. Perspectives in Biology and Medicine, 31(4):526–557, 1988. Qingyun Wang, Doug Downey, Heng Ji, and Tom Hope. SciMON: Scientific inspiration machines optimized for novelty. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 279–299, Bangkok, Thailand, August 2024. Association for Computational Linguistics. doi: 10.18653/v1/2024.acl-long.18. Yixuan Weng, Minjun Zhu, Guangsheng Bao, Hongbo Zhang, Jindong Wang, Yue Zhang, and Linyi Yang. Cycleresearcher: Improving automated research via automated review. 2025. 15 Continuous Knowledge Metabolism Guangzhi Xiong, Eric Xie, Amir Hassan Shariatmadari, Sikun Guo, Stefan Bekiranov, and Aidong Zhang. Improving scientific hypothesis generation with knowledge grounded large language models. arXiv preprint arXiv:2411.02382, 2024. Zonglin Yang, Wanhao Liu, Ben Gao, Yujie Liu, Wei Li, Tong Xie, Lidong Bing, Wanli Ouyang, Erik Cambria, and Dongzhan Zhou. MOOSE-chem2: Exploring LLM limits in fine-grained scientific hypothesis discovery via hierarchical search. In The Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. URL https://openreview.net/forum?id=Q3DtkFJ1Ap. Zonglin Yang et al. Large language models for automated open-domain scientific hypotheses discovery. In Findings of the Association for Computational Linguistics: ACL 2024. Association for Computational Linguistics, 2024. Jiakang Yuan, Xiangchao Yan, Bo Zhang, Tao Chen, Botian Shi, Wanli Ouyang, Yu Qiao, Lei Bai, and Bowen Zhou. Dolphin: moving towards closed-loop auto-research through thinking, practice, and feedback. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 21768–21789, 2025. Yu Zhang, Xiusi Chen, Bowen Jin, Sheng Wang, Shuiwang Ji, Wei Wang, and Jiawei Han. A comprehensive survey of scientific large language models and their applications in scientific discovery. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen, editors, Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 8783–8817, Miami, Florida, USA, November 2024. Association for Computational Linguistics. doi: 10.18653/v1/2024.emnlp-main.498. URLhttps://aclanthology. org/2024.emnlp-main.498/. Yangqiaoyu Zhou, Haokun Liu, Tejes Srivastava, Hongyuan Mei, and Chenhao Tan. Hypothesis generation with large language models. In Lotem Peled-Cohen, Nitay Calderon, Shir Lissak, and Roi Reichart, editors, Proceedings of the 1st Workshop on NLP for Science (NLP4Science), pages 117–139, Miami, FL, USA, November 2024. Association for Computational Linguistics. doi: 10.18653/v1/2024.nlp4science-1.10. URL https://aclanthology.org/2024.nlp4science-1.10/. 16 Continuous Knowledge Metabolism Appendix: Table of Contents A Implementation Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 Models, infrastructure, & reproducibility B Evaluation Topics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 Topic selection procedure & full list C Prompt Templates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 Diff-based Knowledge Update Change Detection (Trigger Classification) Hypothesis Generation D Novelty Judge Scoring Dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 D1–D4 definitions & dimension breakdown E Full Statistical Significance Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 Pairwise Wilcoxon tests & Cohen’s d (all 30 comparisons) F Efficiency Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 Token efficiency table Cost breakdown by phase Scalability considerations G CKM-Lite Exclusive Coverage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 12 topics where only CKM-Lite achieves hits H Temporal Dynamics (Full Results) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 Per-window novelty & hit rate table I CKM-Shuffled Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 8-topic temporal ordering ablation J Hypothesis Evolution Trajectories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 PCA-projected trajectories for 9 topics High-drift / Medium-drift / Stable topic analysis 17 Continuous Knowledge Metabolism Boundary cases & limitations K Detailed Case Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 Case A: Methodological transfer (AI for Hypothesis Gen.) Case B: Gap exploitation (Protein Structure Prediction) L Hypothesis Embedding Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 Embedding visualizations (20 topics) Diversity, peripheral effect, score distribution, nearest-neighbor Per-topic observations (6 representative topics) 18 Continuous Knowledge Metabolism A. Implementation Details Table 6 lists all models used in our experiments, their roles, and configuration details. RoleModelProvider Temp. Knowledge generationGemini-2.5-Flash (Gemini Team, 2025) Google0.7 (init, evolution, hypotheses) Hit verify (pre-filter)GPT-4o-mini (OpenAI, 2024)OpenAI 0.1 Hit verify (re-judge)GPT-4o (OpenAI, 2024)OpenAI 0.1 Keyword extractionGPT-4oOpenAI 0.1 Validation profilesGPT-4oOpenAI 0.1 Novelty judgeGPT-5.4 (OpenAI, 2026)OpenAI 0.1 Embeddingtext-embedding-3-smallOpenAI — Table 6: Models used across all experiments. Generation and evaluation use models from different providers to mitigate self-evaluation bias. Hit verification uses a two-stage pipeline: GPT-4o-mini pre-filters candidates (threshold≥ 5.0), and GPT-4o re-judges those that pass. Infrastructure. All API calls were routed through a single proxy endpoint. Papers were sourced from arXiv with full-text extraction via LaTeX source parsing where available, falling back to abstract-only when full text was inaccessible. Embedding-based candidate pre-filtering used cosine similarity with top-30 candidates per hypothesis. Total compute cost for the full evaluation (4 systems×50 topics) was approximately $150 USD. Reproducibility. All prompts are provided in Appendix C. The 50 evaluation topics and their selection crite- ria are in Appendix B. Code and data will be released athttps://github.com/tsingyuai/scientify. B. Evaluation Topics We evaluate on 50 research topics spanning 8 categories. Topics were selected through the following procedure: (1) we identified major arXiv categories relevant to NLP and AI (cs.CL, cs.AI, cs.LG, cs.CV, cs.IR, cs.SE, q-bio, physics); (2) for each category, we selected active research directions with sufficient paper volume (≥50 papers in 2019–2024) to ensure meaningful knowledge evolution; (3) we balanced the final set across categories to avoid over-representation of any single subfield. No topic was added or removed based on preliminary results. C. Prompt Templates We provide the three core prompts used in CKM’s knowledge metabolism cycle. All prompts are used with Gemini-2.5-Flash at temperature 0.7 (generation) or 0.1 (detection). C.1. Diff-based Knowledge Update The update prompt processes each topic file against new papers. The key design choice is the contrastive instruction that forces explicit categorization of each finding. 19 Continuous Knowledge Metabolism CategoryTopics NLP Core (9)Relation extraction, Aspect-based sentiment analysis, Open-domain QA, Diversity in text generation, Machine translation (low-resource), Speech recognition, Information retrieval, Information extraction, Active learning for NLP LLM Methods (10)Instruction tuning, Prompt engineering & ICL, Chain-of-thought reasoning, Efficient fine- tuning, Knowledge distillation, Model compression, Mixture of experts, RLHF, Long-context, Continual learning LLM Apps (10)Code generation, Tool-using agents, Complex reasoning agents, Hypothesis generation, Software engineering, Security of code LLMs, Automated evaluation, Data filtering, Data augmentation, Synthetic data evaluation Domain (6)Drug discovery, Medical image analysis, Clinical NLP, Protein structure prediction, Recom- mendation systems, Surrogate modeling Safety (5)Adversarial robustness, Fairness & bias, OOD detection, Factual consistency, Explainability Multilingual (3)Multilingual LLMs, Cross-lingual transfer, Federated learning for LMs Multimodal (2)Vision-language models, Visual QA Other (5)Text-to-image generation, Domain adaptation, KG reasoning, Document understanding, UX evaluation Table 7: Complete list of 50 evaluation topics by category. System Prompt You are a research knowledge base updater. Update a single topic file based on new papers. Work strictly from the provided content –- no hallucination. Cite arxiv_id for every claim. User Prompt (abbreviated) Topic file: topic_name Period: period Current content: topic_content New papers this period: papers_text Update this topic file by applying these principles to EACH new paper: - New finding/method absent from current content→ append with citation - Confirms an existing conclusion→ add new evidence, note increased confidence - Contradicts or refines→ rewrite, preserve both positions, mark [Revised] - Cross-domain connection→ note the link Size control: keep the file under 200 lines. When approaching the limit, compress older entries. C.2. Change Detection (Trigger Classification) This prompt compares knowledge states before and after a window’s update to produce a structured trigger signal. 20 Continuous Knowledge Metabolism System Prompt You are a research intelligence analyst monitoring a knowledge base for changes. Characterize what changed in the latest update –- what type of change occurred and what is notable. Be factual and precise. Return valid JSON only. User Prompt (abbreviated) Knowledge State BEFORE: knowledge_before Knowledge State AFTER: knowledge_after New Papers (period): papers_text Classify the change type: 1. INCREMENTAL: routine extension 2. CONTRADICTION: revises established finding 3. CONVERGENCE: independent papers point to same conclusion 4. BRIDGE: connection between previously unrelated topics 5. TREND_CONFIRMED: earlier pattern validated Return JSON: change_type, reason, key_changes C.3. Hypothesis Generation The hypothesis prompt receives the full evolution context and generates structured, testable hypotheses. User Prompt (abbreviated) You have been tracking "topic" for n_windows periods. Knowledge Evolution Trajectory: evolution_trajectory Current Knowledge State: knowledge New Papers (period): papers_text What Changed: trigger_type –- trigger_reason Previously Generated Hypotheses: existing_hypotheses Generate hypotheses that are conceptually novel. Consider: 1. Contradictions & tensions 2. Non-obvious bridges 3. Trend extrapolation 4. Gap exploitation 5. Cross-paper synthesis (must yield insight beyond sum of parts) Avoid hypotheses that merely integrate existing methods. Each hypothesis must include: statement, research claim (problem, method delta, baseline, expected observable, evaluation plan, failure mode), reasoning, source papers, trigger, self-assessment. D. Novelty Judge Scoring Dimensions The novelty judge (independent LLM from a different provider than the generation model) scores each hypothesis blind on four dimensions using a 1–10 scale. The judge receives only the hypothesis content and is not informed of which system generated it. •D1 — Conceptual Originality: Does the hypothesis propose a genuinely new idea, or is it a predictable 21 Continuous Knowledge Metabolism combination of existing methods? A score of 1–3 indicates incremental extension of a single paper; 4–6 indicates a novel combination or non-obvious application; 7–10 indicates a fundamentally new concept or surprising cross-domain connection. • D2 — Cross-field Synthesis: Does the hypothesis meaningfully bridge distinct research areas? A score of 1–3 indicates single-subfield work; 4–6 indicates connecting related but distinct areas (e.g., NLP + IR); 7–10 indicates bridging fundamentally different domains (e.g., language evolution theory + LLM agent design). • D3 — Gap Identification Precision: Is the research gap clearly articulated and supported by specific evidence? A high score requires naming specific prior works, identifying what they cannot do, and explaining why the gap matters for the field. Vague claims like “improve efficiency” without specifying whose efficiency, on what task, score low. • D4 — Specificity & Falsifiability: Does the hypothesis specify a concrete task, baseline, expected outcome, and evaluation plan? A score of 8+ requires naming specific models (e.g., “LLaMA-2-7B”), datasets (e.g., “Davis, KIBA”), metrics (e.g., “RMSE, Hits@1”), expected effect sizes (e.g., “at least 5% improvement”), and explicit failure conditions. Dimension breakdown across systems. The largest gap between CKM-Full and CKM-Lite is on D1 Originality (+1.44 points), followed by D2 Cross-field (+0.84). D3 Gap Precision shows a smaller difference (+0.38), and D4 Falsifiability is nearly identical across all systems (∼7.8), confirming that specificity is driven by the shared prompt template rather than the knowledge processing mechanism. This pattern suggests that CKM-Full’s change-aware instrumentation is primarily associated with what ideas are generated (originality, cross-domain connections), not how precisely they are specified. E. Full Statistical Significance Tests Table 8 reports pairwise Wilcoxon signed-rank tests and Cohen’sdeffect sizes for all system pairs across five metrics, computed on the 50 common topics. Significance levels: ∗ p < 0.001, ∗ p < 0.01, ∗ p < 0.05. All reportedp-values in the main text are drawn from this table. We use the Wilcoxon signed-rank test (non- parametric, paired) rather than a pairedt-test because hypothesis-level metrics are not normally distributed across topics. Cohen’sdis computed as the mean difference divided by the pooled standard deviation; values > 0.8 are conventionally considered large effects. Key patterns. CKM-Lite’s hit rate advantage over Batch is significant (p = .006) with a medium effect size (d = 0.55). CKM-Full’s novelty advantage over all systems is highly significant (p < .001) with very large effect sizes (d > 2.4), but this comes at a significant cost in hit rate (p < .001vs. Lite,p = .034vs. Batch). Cross-domain scores show no significant differences between most pairs, suggesting that interdisciplinary breadth is not substantially affected by the choice of processing pipeline. Yield is significantly higher for all incremental systems vs. Batch (d > 1.4,p < .001), confirming that temporal decomposition produces more hypotheses regardless of other design choices. 22 Continuous Knowledge Metabolism ComparisonMetricMean A Mean B∆dp CKM-Lite vs BatchHit Rate5.793.03+2.76 +0.55.006 ∗ Novelty6.036.39 −0.36 −1.42 <.001 ∗ Best Match3.903.46+0.43 +0.58 <.001 ∗ Yield17.2613.68+3.58 +1.78 <.001 ∗ Cross-dom.1.331.40 −0.07 −0.09.758 CKM-Lite vs CKM-Full Hit Rate5.791.39+4.40 +1.14 <.001 ∗ Novelty6.036.82 −0.79 −3.46 <.001 ∗ Best Match3.904.58 −0.68 −2.25 <.001 ∗ Yield17.2617.84 −0.58 −0.24.818 Cross-dom.1.331.42 −0.09 −0.12.126 CKM-Lite vs AbstractHit Rate5.794.00+1.80 +0.34.042 ∗ Novelty6.036.25 −0.22 −0.91 <.001 ∗ Best Match3.903.48+0.42 +0.52.010 ∗ Yield17.2617.16+0.10 +0.08.333 Cross-dom.1.331.42 −0.09 −0.12.027 ∗ CKM-Full vs BatchHit Rate1.393.03 −1.64 −0.39.034 ∗ Novelty6.826.39+0.43 +2.44 <.001 ∗ Best Match4.583.46+1.12 +1.57 <.001 ∗ Yield17.8413.68+4.16 +1.43 <.001 ∗ Cross-dom.1.421.40+0.01 +0.02.490 CKM-Full vs AbstractHit Rate1.394.00 −2.61 −0.59.003 ∗ Novelty6.826.25+0.57 +3.62 <.001 ∗ Best Match4.583.48+1.11 +1.41 <.001 ∗ Yield17.8417.16+0.68 +0.28.840 Cross-dom.1.421.42 −0.00 −0.00.910 Abstract vs BatchHit Rate4.003.03+0.97 +0.18.516 Novelty6.256.39 −0.14 −0.72 <.001 ∗ Best Match3.483.46+0.01 +0.01.173 Yield17.1613.68+3.48 +1.74 <.001 ∗ Cross-dom.1.421.40+0.01 +0.02.682 Table 8: Full pairwise statistical comparisons across all system pairs and metrics (Wilcoxon signed-rank test, 50 topics). ∗ p < 0.001, ∗ p < 0.01, ∗ p < 0.05. SystemTokens Tok/Hyp Tok/Hit Hyps CKM-Lite26M30K400K863 CKM-Abstract 76M89K2.2M858 CKM-Full293M329K22.6M892 Batch297M434K11.4M684 Table 9: Token efficiency. CKM-Lite achieves the lowest cost per hypothesis (30K tokens) and per hit (400K tokens), using 92% fewer tokens than Batch Baseline while generating more hypotheses and more hits. F. Efficiency Analysis CKM-Lite’s efficiency advantage comes from two sources: (1) incremental windows process 8 papers per window rather than all papers at once, reducing per-call prompt length; (2) no diff categorization, change detection, or trajectory conditioning overhead. CKM-Full and Batch Baseline consume similar total to- kens (∼295M) but for different reasons: CKM-Full runs many smaller calls across 6 windows with rich instrumentation, while Batch runs fewer but longer calls with all papers in a single prompt. 23 Continuous Knowledge Metabolism Cost breakdown by phase. For all systems, the majority of tokens are consumed during Phase 2 (evolu- tion/generation). CKM-Lite’s Phase 2 uses Gemini-2.5-Flash for generation (∼$0.50 per experiment) and GPT-4o/GPT-4o-mini for Phase 3 validation (∼$30 per experiment via two-stage judging). CKM-Full’s Phase 2 is approximately 11×more expensive per hypothesis due to the additional diff-update, trigger-detection, and trajectory-conditioning calls. The total cost for our full evaluation (4 systems×50 topics) was approximately $150, with CKM-Lite accounting for less than 10% of this budget. Scalability considerations. CKM-Lite’s low token cost makes it practical for continuous deployment: monitoring 100 research topics with monthly updates would require approximately 5M tokens per cycle (∼$0.75 with current pricing), well within the free tier of most LLM providers for the generation phase. The primary cost bottleneck is Phase 3 validation, which scales linearly with the number of hypotheses and candidate validation papers. G. CKM-Lite Exclusive Coverage CKM-Lite achieves predictive hits on 12 topics where no other system (CKM-Full, Batch Baseline, or CKM- Abstract) achieves any hit. These topics span diverse areas including NLP core tasks (cross-lingual transfer, low-resource MT, open-domain QA), LLM methods (prompt engineering, RLHF, continual learning), and multimodal learning (vision-language models, text-to-image generation). TopicHit Rate Yield Conversational information retrieval11.8%17 Reinforcement learning from human feedback6.7%15 Cross-lingual transfer for low-resource langs5.9%17 Open-domain QA with language models5.9%17 Vision-language models & multimodal learning5.9%17 Active learning for NLP5.9%17 Low-resource machine translation5.6%18 Code generation with LLMs5.6%18 Prompt engineering & in-context learning5.6%18 Text-to-image generation & diffusion models5.6%18 Data filtering for domain-specific models5.6%18 Continual learning & catastrophic forgetting5.6%18 Table 10: 12 topics where only CKM-Lite achieves predictive hits. These represent research directions that incremental accumulation can anticipate but batch processing and change-aware instrumentation cannot. The breadth of these exclusive topics spanning from low-resource NLP to generative AI to continual learning suggests that CKM-Lite’s advantage is not limited to a specific domain but reflects a general property of incremental processing: by maintaining a continuously updated knowledge state, the system tracks gradual shifts that batch processing misses, while avoiding the over-specificity that CKM-Full’s diff mechanism introduces. H. Temporal Dynamics (Full Results) Table 11 reports per-window hypothesis characteristics for CKM-Full across the six 2-month evolution windows (Jan–Nov 2024). Two opposing trends are visible: novelty increases monotonically from 6.73 24 Continuous Knowledge Metabolism to 6.86 as the system accumulates more knowledge, while hit rate decreases from 3.5% to 0.0%. Early windows primarily capture “low-hanging fruit”: predictable next steps that are relatively easy to validate. Later windows, by contrast, tend to produce more original hypotheses that move further away from the immediate research frontier. The inter-window semantic shift also increases slightly (cosine distance 0.229→ 0.242), confirming that later hypotheses explore genuinely new directions rather than refining earlier ones. Windown Novelty Hits Hit Rate 2024-01 (Jan) 1446.7353.5% 2024-03 (Mar) 1476.8110.7% 2024-05 (May) 1556.8621.3% 2024-07 (Jul) 1516.8621.3% 2024-09 (Sep) 1486.8421.4% 2024-11 (Nov) 1476.8600.0% Table 11: Hypothesis characteristics by evolution window (CKM-Full). Novelty increases monotonically while hit rate decreases. I. CKM-Shuffled Results (Temporal Ordering Ablation) Due to computational budget constraints, we report CKM-Shuffled (timestamps randomized) on 8 topics. CKM-Shuffled uses the identical pipeline but randomizes all paper dates, destroying temporal ordering while preserving the same papers and processing mechanisms. On these 8 topics, CKM-Shuffled achieves a hit rate (2.2%) comparable to CKM (1.6%), suggesting that temporal ordering has limited impact on predictive coverage in this sample. However, CKM-Shuffled’s average novelty (6.38) is noticeably lower than CKM’s (6.82), consistent with the hypothesis that correct temporal ordering helps CKM identify more original research directions even when it does not improve hit rate. We acknowledge that 8 topics is insufficient for strong conclusions and that full-scale validation of this ablation is needed. TopicShuffled CKM CKM-Lite Low-resource speech5.6%7.1%7.1% Cross-lingual transfer0.0%0.0%5.9% Low-resource MT0.0%0.0%5.6% Multilingual LLMs0.0%5.6%0.0% Security of code LLMs 11.8% 5.6%5.6% Automated evaluation0.0%0.0%5.6% Code generation0.0%0.0%5.6% AI for software eng.0.0%0.0%5.6% Average2.2%2.3%5.1% Table 12: Hit rate for CKM-Shuffled on 8 topics. Average novelty: Shuffled 6.38, CKM 6.82. J. Hypothesis Evolution Trajectories A unique property of CKM is that it generates hypotheses incrementally across time windows, producing a traceable evolution trajectory for each topic. In the main text (§5.2), we show that trajectory stability (drift) 25 Continuous Knowledge Metabolism is marginally associated with predictive success (r = −0.281,p = 0.051). Here we visualize these trajectories to provide qualitative intuition for how different drift levels manifest in practice. For each topic, we compute the centroid of hypothesis embeddings within each window and project all centroids into 2D via PCA. Figure 4 shows trajectories for 9 representative topics: 3 high-drift (top), 3 medium (middle), 3 stable (bottom). Arrows show how hypothesis focus shifts from Jan (blue) to Nov (red) 2024. Background dots show individual hypotheses. d=0.445 High Drift Continual Learning And Catastrophic Forgetting d=0.401 Knowledge Distillation For Small Language Models d=0.385 Adversarial Robustness Of Deep Learning Models d=0.274 Medium Drift Mixture Of Experts Routing For Language Models d=0.267 Model Compression And Pruning For Neural Networks d=0.266 Efficient Fine Tuning Of Large Language Models d=0.185 Stable Ai For Hypothesis Generation In Science d=0.174 Complex Reasoning For Ai Agents d=0.152 Low Resource Speech Recognition JanMarMayJulSepNov Figure 4: Hypothesis evolution trajectories for 9 topics. Arrows colored from blue (Jan) to red (Nov). Drift values shown in bottom-right badges. 26 Continuous Knowledge Metabolism High-drift topics (top row). Continual Learning (drift=0.445, highest of all 50 topics) exhibits the longest trajectory, with each window pulling hypothesis focus in a different direction from regularization-based approaches (Jan) to replay methods (May) to architecture-based solutions (Nov). CKM-Full generates 0 hits on this topic, consistent with the meta-signal finding (§5.5). Knowledge Distillation (drift=0.401) is the exception: despite high drift, CKM-Full achieves one hit (5.9%). Closer inspection shows that the hit hypothesis was generated in the September 2024 window, by which point the trajectory had already stabilized around self-correction methods. The hit therefore reflects late-stage convergence rather than an early prediction. Adversarial Robustness (drift=0.385) shows a characteristic “zigzag” pattern where the field oscillates between attack methods and defense methods across windows, making directional prediction difficult. Medium-drift topics (middle row). These topics show more coherent trajectories with occasional direc- tional shifts. The trajectory paths are shorter and smoother, with adjacent windows’ centroids remaining relatively close. None of the three medium-drift topics shown produced CKM-Full hits, but their best match scores (4.3–4.7) are comparable to the overall CKM-Full average, indicating alignment without crossing the hit threshold. Stable topics (bottom row). Low-Resource Speech Recognition (drift=0.152, lowest of all topics) shows an almost stationary trajectory: hypothesis focus barely shifts across six windows, reflecting a field with a clear, cumulative research direction. CKM-Full achieves 7.1% hit rate here. AI for Hypothesis Generation (drift=0.185) and Complex Reasoning for AI Agents (drift=0.174) also show compact trajectories, though only the former produces a hit. Boundary cases and limitations. The trajectory analysis has notable edge cases. Federated Learning (drift=0.260, near median) achieves a CKM-Full hit despite moderate drift, suggesting that the drift threshold is not sharp. Conversely, Explainability (drift=0.219, low) produces 0 hits despite a stable trajectory, indicating that low drift is necessary but not sufficient for predictive success. The field must also have addressable gaps that CKM’s knowledge metabolism can detect. These boundary cases motivate the drift×trigger interaction analysis in §5.5, which provides a more nuanced predictor than drift alone. Summary. The trajectory visualizations reveal three distinct patterns: (1) wandering trajectories (high drift) where each window pulls hypothesis focus in a new direction, making prediction difficult; (2) coherent trajectories (low drift) where hypotheses build cumulatively on a stable research direction, enabling targeted gap exploitation; and (3) convergent trajectories where initially wandering paths stabilize in later windows, occasionally producing late-stage hits (as in Knowledge Distillation). These patterns provide qualitative support for the quantitative finding that CKM-Full is most effective in fields with coherent developmental trajectories, and suggest that trajectory monitoring could serve as a practical signal for when to trust system-generated hypotheses. K. Detailed Case Studies We present two cases where CKM achieved exclusive hits (CKM-Lite scored 0% on the same topic), illustrating how diff-based knowledge metabolism can produce genuinely insightful predictions. The main text (§5.1) presents a contrasting case where CKM-Lite hit and CKM did not. 27 Continuous Knowledge Metabolism Case A: Methodological transfer across domains — AI for Hypothesis Generation (drift = 0.185) CKM-FullTrigger: Bridge ⋅ Jan 2024 ⋅ Hit ✓ Lead: 639 days “Implementing a Test-Driven Development (TDD)-inspired framework for scientific hypothesis generation, where ‘scientific tests’ (e.g., formal logical consistency checks, computable simulations, or statistical tests on existing datasets) are defined prior to or concurrently with LLM-based generation, will reduce the number of hypotheses requiring costly experimental validation by at least 30%.” Matched→HARPA: A Testability-Driven, Literature-Grounded Framework for Research Ideation [2510.00620] Judge: 6.5/10 CKM-LiteNo trigger ⋅ Jan 2024 ⋅ No Hit Best match: 4.5/10 “Integrating human attention patterns into the self-attention mechanisms of large language models for scientific hypothesis generation will yield hypotheses that are more interpretable and understandable by human scientists, as evidenced by ‘Attention Simplification’ in the model’s internal attention maps.” Analysis of Case A (AI for Hypothesis Generation). This is CKM-Full at its best. The Bridge trigger connected two previously unrelated literatures—software engineering methodology (TDD) and scientific hypothesis generation, which producing a prediction that a testability-first framework would improve hypothesis quality. 639 days later, HARPA was published with exactly this approach: testability-driven, literature-grounded research ideation. CKM-Lite’s hypothesis from the same window proposes integrating “human attention patterns” into LLM self-attention—a superficially novel idea that is mechanistically vague and did not correspond to any future work (best match 4.5/10). The key difference: CKM-Full transfers a concrete methodology (TDD) across domains, while CKM-Lite combines abstract concepts without a clear mechanism. This is the type of insight that diff-based knowledge metabolism can produce by tracking what changed in both the hypothesis generation literature and the software engineering literature, CKM-Full identified a bridging opportunity that accumulation-based approaches missed. Case B: Precise gap exploitation — Protein Structure Prediction (drift = 0.209) CKM-FullTrigger: Gap_Exploitation ⋅ Jan 2024 ⋅ Hit ✓ Lead: 462 days “Extending the PSC-CPI multi-scale contrasting framework to explicitly incorporate atomic-scale representations and contrasting objectives will achieve at least a 5% improvement in RMSE for CPI binding affinity prediction and a 10% improvement in atomic contact prediction accuracy on the ‘Unseen-Both’ setting of the Davis, KIBA, and Mert datasets, compared to the residue-level PSC-CPI.” Matched→ EquiCPI: SE(3)-Equivariant Geometric Deep Learning for Structure-Aware CPI Prediction [2504.04654] Judge: 6.5/10 CKM-LiteNo trigger ⋅ Jan 2024 ⋅ No Hit Best match: 2.8/10 “Integrating general-purpose RNA Language Model (RiNALMo) embeddings into protein tertiary structure prediction architectures, adapted for RNA, will enable accurate and significantly faster de novo RNA tertiary structure prediction compared to methods relying on traditional Multiple Sequence Alignment.” Analysis of Case B (Protein Structure Prediction). This case illustrates the Gap_Exploitation trigger: the combination that achieves 33.3% hit rate in low-drift topics (§5.5). CKM-Full identified a specific, 28 Continuous Knowledge Metabolism addressable gap in the PSC-CPI framework (residue-level only, no atomic scale) and predicted that adding atomic-scale representations would yield measurable improvements on named benchmarks. 462 days later, EquiCPI was published, introducing an SE(3)-equivariant geometric deep learning approach for structure- aware CPI prediction and offering a different technical solution to the same underlying gap. CKM-Lite’s hypothesis from the same window proposes adapting RNA language models for protein structure: a cross- modality transfer that sounds plausible but addresses no specific gap (best match 2.8/10). Cross-case pattern. Both cases share a pattern: in low-drift fields with clear knowledge gaps, CKM-Full’s diff mechanism identifies what is missing rather than what is trending, producing hypotheses specific enough to be validated by future work addressing the same gap. L. Hypothesis Embedding Space We embed each hypothesis statement usingtext-embedding-3-small(1,536 dim) and analyze 3,155 scored hypotheses via t-SNE (perplexity=30, seed=42). Aggregate findings. We compute four quantitative metrics in the original 1,536-dimensional space (not the t-SNE projection): 1.Hypothesis Diversity: CKM-Full hypotheses exhibit the highest intra-topic semantic diversity (avg pairwise cosine distance = 0.499), compared to CKM-Lite (0.472), Batch (0.467), and CKM-Abstract (0.465). CKM-Full produces more diverse hypotheses in 16 out of 20 analyzed topics, suggesting that the change-aware pipeline encourages exploration of a broader hypothesis space. 2.Peripheral Hypotheses Align Better: Distance from the topic centroid correlates negatively with best match score (Pearsonr = −0.235,p < 0.0001): hypotheses that are more semantically unique tend to receive higher alignment scores with future papers. Hit hypotheses are consistently farther from the topic centroid than misses across all four experimental groups, with CKM-Full showing the largest hit–miss centroid separation (0.236 vs. CKM-Lite’s 0.117). 3.Score Concentration: 93.1% of CKM-Full scores fall in the 4–5 best match score band (highest concen- tration), compared to 34.2% for CKM-Lite. CKM-Lite’s broader distribution (11.2% below 3, 5.8% above 6) reflects a high-variance strategy that trades consistency for occasional high-scoring hits. 4.Nearest-Neighbor Independence: The average Spearman correlation between a hypothesis’s score and that of its nearest semantic neighbor is weak (r = 0.153). This indicates that alignment with future work is driven less by general topical positioning than by the specificity of the research claim, especially the particular combination of method, setting, and expected outcome. Consequently, two hypotheses within the same sub-topic may still receive very different alignment scores. Per-topic observations. We highlight six topics that illustrate different embedding patterns: Protein Structure Prediction (CKM-Full exclusive hit, low drift). CKM-Full’s hypotheses form a tight cluster in the upper region, with the single hit hypothesis at the cluster periphery. CKM-Lite’s hypotheses overlap substantially with CKM’s but are slightly more dispersed, with no hits. This pattern is consistent with the meta-signal finding: in this stable field, CKM’s focused exploration identifies a precise gap that CKM-Lite’s broader coverage misses. 29 Continuous Knowledge Metabolism Ai For Hypothesis Generation In ScienceAspect Based Sentiment AnalysisClinical Natural Language Processing Federated Learning For Language Models Knowledge Distillation For Small Language Models Knowledge Graph Completion And Reasoning Long Context Understanding In Large Language ModelsLow Resource Speech RecognitionMultilingual Large Language Models Protein Structure Prediction With Deep Learning Synthetic Data Quality Evaluation Tool Using Agents With Large Language Models Out Of Distribution Detection For Language ModelsConversational Information Retrieval User Experience Evaluation For Large Language Models Deep Learning For Medical Image Analysis Deep Learning For Recommendation Systems Document Understanding And Information Extraction Domain Adaptation And Transfer Learning Diversity Evaluation In Text Generation CKM-Full CKM-Full HIT CKM-Lite CKM-Lite HIT Batch Baseline Batch Baseline HIT CKM-Abstract CKM-Abstract HIT Figure 5: Hypothesis embeddings for 20 topics, colored by experiment. Stars = hits. AI for Hypothesis Generation (CKM-Full exclusive hit, lowest drift 0.185). The tightest cluster of any topic. All four systems generate hypotheses in a compact semantic region, reflecting the narrow scope of this field. CKM-Full’s hit (the TDD→hypothesis generation connection) appears as an outlier point, notably distant from the main cluster, illustrating the “peripheral hypotheses predict better” finding at the individual topic level. Tool-Using Agents (CKM-Lite dominant, high drift). The most dispersed embedding pattern of the six topics. CKM and CKM-Lite hypotheses occupy different sub-regions with minimal overlap, suggesting the two systems explore fundamentally different parts of the hypothesis space. CKM-Lite’s three hits are scattered across the space, consistent with its broad-coverage strategy. 30 Continuous Knowledge Metabolism Ai For Hypothesis Generation In ScienceAspect Based Sentiment AnalysisClinical Natural Language Processing Federated Learning For Language Models Knowledge Distillation For Small Language Models Knowledge Graph Completion And Reasoning Long Context Understanding In Large Language ModelsLow Resource Speech RecognitionMultilingual Large Language Models Protein Structure Prediction With Deep Learning Synthetic Data Quality Evaluation Tool Using Agents With Large Language Models Out Of Distribution Detection For Language ModelsConversational Information Retrieval User Experience Evaluation For Large Language Models Deep Learning For Medical Image Analysis Deep Learning For Recommendation Systems Document Understanding And Information Extraction Domain Adaptation And Transfer Learning Diversity Evaluation In Text Generation 01234567 Best Match Score Figure 6: Same 20 topics, colored by best match score (viridis, 0–7). High scores do not cluster spatially. Synthetic Data Quality Evaluation (CKM-Lite dominant, largest hit rate gap). CKM-Lite hypotheses are broadly spread with three hits in different regions. CKM hypotheses form a tighter cluster, but that cluster lies in a region with no hits. This pattern suggests that the diff mechanism may have narrowed attention too strongly toward the simulation-to-real domain gap, whereas CKM-Lite preserved broader and more even coverage of the field. Deep Learning for Medical Image Analysis (neither system hits). Both systems’ hypotheses cluster tightly with high overlap, and no hits appear from any system. The absence of hits despite reasonable best match scores (avg∼4.3) suggests this topic’s future papers may have diverged from directions predictable from the 2024 literature. 31 Continuous Knowledge Metabolism Long-Context Understanding (both systems hit, CKM’s highest hit rate topic at 11.1%). The richest embed- ding pattern: four distinct sub-clusters corresponding to different research threads (memory architectures, evaluation benchmarks, video understanding, retrieval augmentation). Hits for both CKM and CKM-Lite are distributed across clusters, but CKM’s hits more often lie at the boundaries between clusters, consistent with its comparative advantage in bridging distinct knowledge threads. 32

Instant research discovery