InternAgent-1.5: A Unified Agentic Framework for Long-Horizon Autonomous Scientific Discovery

Motivation

• Existing AI-for-science systems are limited by domain-specific architectures, uneven support for dry-lab and wet-lab work, trajectory-local optimization, and insufficient persistent memory for extended research cycles.These limitations hinder unified cross-disciplinary reasoning, integration across broader searches, iterative refinement, and long-term autonomous operation.Source: §1 · pp. 3–4

Discovery domains

• The paper distinguishes Algorithm Discovery, which transforms objectives into formal-system solutions, from Empirical Discovery, which generalizes from observations about the physical world.A framework for both requires unified architecture, strong foundational capabilities, long-horizon iterative optimization, and operation across computational and experimental environments.Source: §1 · p. 4

Architecture

• Its unified architecture coordinates Generation, Verification, and Evolution to support hypothesis construction, methodological evaluation, and evidence-driven refinement across computational and empirical tasks.Deep research supports Generation, solution refinement supports Verification, and long-horizon memory supports Evolution.Source: §1 · pp. 4–5

Results

• InternAgent-1.5 achieves leading performance on HLE, GAIA, GPQA, FrontierScience, and SGI bench.The paper attributes this agentic reasoning performance to foundational capabilities supporting the Generation and Verification subsystems.Source: §1 · p. 5

Scientific discovery tasks

• The system produces competitive algorithmic solutions and executes complete empirical workflows that yield scientific insights in fields including biology and earth sciences.Algorithmic examples include reinforcement learning and test-time methodology; empirical work includes data-oriented scientific tasks.Source: §1 · p. 5

Long-horizon operation

• Structured memory and iterative optimization support accumulated context, sustained hypothesis refinement, and stable improvement of methodological plans across extended discovery cycles.The optimization process centers on the Verification and Evolution subsystems.Source: §1 · p. 5

2.1. System Overview; 2.1.2. Foundational Capabilities

• Deep research, solution refinement, and long-horizon memory support hypothesis construction, methodological evaluation, and persistent context across discovery cycles.These capabilities underpin the system’s continuity, adaptability, and scalability.Source: §2.1 · p. 5, §2.1.2 · p. 7

2.1.1. Architecture

• Generation formulates hypotheses and plans, Verification evaluates them computationally or empirically, and Evolution uses evidence to update knowledge and guide subsequent cycles.Together, the three subsystems maintain an iterative flow of information for multi-cycle autonomous operation.Source: §2.1.1 · pp. 6–7

2.2. Cross Disciplinary Graph Construction and Knowledge Capturing; 2.2.1. Cross-Disciplinary Knowledge Graph

• A heterogeneous cross-disciplinary knowledge graph links documents, concepts, methods, datasets, settings, and problems; graph search and dense retrieval combine into ranked, path-structured evidence.The graph captures typed relations and cross-field dependencies, while tool-assisted parsing and extraction support knowledge integration.Source: §2.2 · p. 7, §2.2.1 · pp. 7–8

2.2.2. Flow Graph

• Deep research represents tasks and subtasks as a dependency-aware DAG that expands incrementally, propagating context and adapting as new evidence emerges.Nodes track task type, content, execution state, and resulting context; directed edges encode dependencies such as required results or supporting evidence.Source: §2.2.2 · p. 9

2.2.3. Graph-Guided Output Synthesis

• Graph-guided synthesis executes nodes when dependencies are satisfied, propagates their contexts, and combines direct, search-augmented, and self-driven responses into a final answer.Knowledge collection draws on multiple domains, scientific tools, and remote resources accessed through SCP.Source: §2.2.3 · pp. 9–10

2.3. Experiment Execution and Multi-round Parallel Optimization; 2.3.1. Generative Design for Experimental Optimization

• To avoid isolated search paths, unused history, and limited idea composition, InternAgent-1.5 replaces rigid tree search with Graph-Augmented Monte Carlo Search over a dynamic solution graph.The framework targets large design spaces in computational algorithm design and physical wet-lab experimentation.Source: §2.3 · p. 10, §2.3.1 · p. 10

2.3.1. Generative Design for Experimental Optimization

• Four graph-based operators generate proposals from parent designs, branch histories, other branches, or combinations of top-performing designs, whose evaluation scores guide later exploration.The search uses selection, expansion, simulation, and backpropagation; proposals run in a simulator or physical experiment, and scores backpropagate through their ancestral paths.Source: §2.3.1 · p. 11

2.3.2. Scenario

• Algorithm-discovery proposals are executed on benchmark datasets, while empirical proposals are simulated or run in laboratories; quantitative evaluations are returned for iterative refinement.Algorithm metrics include accuracy, runtime, and resource use; empirical measurements can include reaction yield, protein stability, fluorescence intensity, or assay signal quality, with SCP coordinating automated devices.Source: §2.3.2 · p. 11

2.4. Structured Cognitive Memory for Long Horizon Scientific Discovery

• Structured Cognitive Memory maintains continuity across discovery cycles by combining procedural, episodic, and semantic information for short-term refinement, mid-term adaptation, and long-term conceptual development.It is engaged throughout the discovery loop.Source: §2.4 · pp. 11–12

2.4.1. Strategy-Procedural Memory

• Strategy Procedural Memory distills reusable patterns and failure lessons from past reasoning workflows into strategic priors that guide coherent plans and help avoid redundant steps and tool calls.It supports deep research by integrating literature, synthesizing context, constructing multi-step plans, and diagnosing failed strategies.Source: §2.4.1 · p. 12

2.4.2. Task-Episodic Memory

• Task Episodic Memory stores each experiment’s method, metrics, and improvement judgment, then retrieves relevant episodes to avoid unsuccessful directions, exploit successful ones, and refine hypotheses.Episodes use semantic embeddings and sparse lexical features to support conceptual and literal alignment.Source: §2.4.2 · p. 13

2.4.3. Semantic-Knowledge Memory

• Semantic Knowledge Memory combines a Long Term Experience Library of distilled insights with an Idea Graph of explored directions to support long-horizon research evolution and continued exploration.After experimental batches, pairwise method combinations and contrastive comparison of methods and results yield methodological principles and experimental heuristics; novelty scoring favors objectives beyond previously explored conceptual regions.Source: §2.4.3 · p. 13

3. Experiments

• Mechanism-discovery applications span Earth, life, and biological sciences, as well as physical science tasks.The evaluation also includes cross-disciplinary benchmarks and autonomous algorithm development.Source: §3 · p. 13

3.1.1. General Scientific Reasoning Abilities

• SGI-Bench covers four task families across 10 disciplines and 1,000+ expert-curated samples; results use its DeepResearch and IdeaGeneration subsets.The families are Scientific Deep Research, Idea Generation, Dry/Wet Experiments, and Experimental Reasoning.Source: §3.1.1 · pp. 13–14
• Evaluation also covers GAIA’s 165-question validation set, GPQA-diamond’s 198 questions, HLE’s 2,500 expert-written questions, and FrontierScience’s Olympiad and Research subsets.FrontierScience results are averaged over 20 runs on Olympiad and 30 runs on Research, using the standard evaluation set.Source: §3.1.1 · p. 14

3.1.2. Algorithm Discovery

• Six scientific-data algorithm-discovery tasks assess InternAgent-1.5 against baselines, contrasting its original-codebase optimization workflow with InternAgent-1.0’s manually consolidated single-file repositories.Tasks use AutoRYP’s Suzuki–Miyaura dataset and R2, AutoTPPR’s Perturb-seq dataset and Top-20 DE MSE, AutoPower’s IEEE 39-Bus benchmark and PQ-node RMSE, AutoTSF’s ETTh1 dataset and MAE averaged over horizons 96, 192, 336, and 720, AutoMD’s MD17 dataset and Force-MAE, and AutoEAP’s UMI-STARR-seq dataset and HK-PCC.Source: §3.1.2 · p. 14
• AI algorithm discovery covers model training, memory optimization, reinforcement learning, and data processing, using papers accepted by top AI conferences in 2025 as comparative baselines.Tasks include AutoTTS with Atom [47], AutoMem with A-MEM [48], AutoLM’s self-distillation pipeline for training Intern-S1-mini [49], and AutoTTRL against TTRL [51] on AIME 2024.Source: §3.1.2 · p. 15

3.1.3. Empirical Discovery

• Earth-science tasks test climate diagnostics and temperature downscaling, including analysis of historical model and observational data and designing methods to improve coarse-resolution fields.Diagnostics use 1970–2010 TAS from 20 CMIP6 models and ERA5 to identify warming trends and regional biases; downscaling maps NCEP-NCAR-R1 data from 2° to ERA5’s 0.25° resolution, against Kriging and BCSD.Source: §3.1.3 · p. 15
• Life-science tasks assess early-stage therapeutic target discovery through evidence synthesis and iterative, multimodal target prioritization.The tasks reproduce GPR160 as an HCC target using resources including TCGA, OpenTargets, and KEGG, and identify ARG2 as a mechanistically grounded CRC target with experiment-ready validation suggestions.Source: §3.1.3 · p. 16
• The biological-sciences task combines fluorescent-protein sequence and structure analysis, candidate design, wet-lab execution, and feedback from measured performance.Design uses ESMFold and predictors such as ProSST; SCP coordinates DNA assembly, expression, purification, and fluorescence measurement.Source: §3.1.3 · p. 16
• Physical-science tasks evaluate reaction outcome prediction and generative scaffold hopping, requiring atomic conservation or preservation of molecular shape and electrostatic features.The reaction task uses ChemCoTBench [62] and a corrected ground truth for 26 problematic entries; scaffold candidates are filtered using medicinal-chemistry metrics including Synthetic Accessibility and LogP.Source: §3.1.3 · p. 16

SGI-Bench

• 37.74% on Deep Research and 58.11% on Idea Generation give InternAgent-1.5 the best performance on both SGI-Bench tracks.It exceeds Gemini-3-pro’s 18.48% on Deep Research and GPT-5’s 55.40% on Idea Generation; its iterative workflow combines planning, information gathering, and refinement.Source: §3.2 · pp. 16–17, Table 2 · p. 17

HLE

• 40.87% text-only and 40.00% full-benchmark accuracy make InternAgent-1.5 the best overall performer among the compared HLE systems.Its improvements are consistent across most HLE sub-domains, indicating robustness on long-horizon, cross-disciplinary reasoning tasks.Source: Table 3 · p. 17, §3.2 · pp. 17–18, Table 4 · p. 18

GAIA

• InternAgent-1.5 outperforms Manus, Mirothinker, and Tongyi-DR on GAIA, including 61.54% robustness on Level 3 questions.The passage attributes its multi-hop and compositional reasoning performance to an iterative workflow of knowledge planning, collection, and refinement.Source: Table 3 · p. 17, §3.2 · p. 17

FrontierScience

• 77.20% Olympiad and 12.00% Research are InternAgent-1.5’s best overall FrontierScience results, ahead of all baselines in both settings.It shows particularly strong performance in Chemistry and Physics; the reported comparisons include DeepSeek-V3.2-Thinking at 64.70% Olympiad and Mirothinker-v1.5 at 7.50% Research.Source: Table 5 · p. 18, §3.2 · p. 18

GPQA

• 87.37% average accuracy on GPQA-diamond gives InternAgent-1.5 state-of-the-art performance against base models and tool-augmented agents.The passage notes particularly strong results in Chemistry and Physics.Source: §3.2 · p. 18

3.3.1. Scientific Algorithm

• R2 reaches 36.6 on Suzuki-Miyaura AutoRYP, versus 27.6 for LoRA-finetuned LLaMA3 and 31.8 for Dolphin; molecular-dynamics Energy MAE falls to 0.114 from VisNet’s 0.158.Source: §3.3.1 · pp. 18–19
• InternAgent-1.5 consistently outperforms prior work and state-of-the-art domain-specific baselines across six scientific algorithm-task types.The comparison is summarized in Table 7.Source: §3.3.1 · p. 18, Table 7 · p. 19
• Power-flow RMSE is 0.00318 on IEEE 39-Bus versus SenseFlow’s 0.00473; ETTh1 forecasting MAE is 0.423, below DLinear’s 0.4382.Source: §3.3.1 · p. 19
• Transcription-prediction MSE is 0.143 versus GEARS’ 0.197, while enhancer-activity Pearson Correlation Coefficient reaches 0.91 versus DeepSTARR’s 0.65.Source: §3.3.1 · p. 19

3.3.2. AI Algorithm

• MMLU-CF [81] score reaches 72.5 versus 70.9 with the reasoning architecture proposed by [47].Source: §3.3.2 · p. 20
• Locomo [82] F1 reaches 0.2785 versus the A-MEM [48] baseline’s 0.2338 under AutoMem with Qwen2.5-3B [66].The passage attributes the improvement to more reliable long-horizon retention, retrieval, and integration of accumulated information.Source: §3.3.2 · p. 20
• Across prior-work reinforcement-learning control and decision benchmarks, AutoRL yields higher returns and more stable training than TTRL.The passage attributes these results to trajectory refinement and decision-making guidance across diverse settings.Source: §3.3.2 · p. 20
• MATH500 accuracy rises from 0.880 to 0.904 after fine-tuning Intern-S1-mini [78] with synthesized mathematical reasoning data at 16k context.The experiment applies the full self-distillation pipeline.Source: §3.3.2 · p. 20

3.4.1. Earth Science

• Climate diagnostics across 20 CMIP6 models produced ranked temperature-trend comparisons and spatial maps showing canonical warming patterns, including enhanced high-latitude trends.The analysis integrated climate literature and physical reasoning to select metrics and build a pipeline for data retrieval, temporal alignment, and statistical estimation.Source: §3.4.1 · p. 21, Figure 7 · p. 21, Figure 8 · p. 22
• RMSE fell to 0.8488 from 3.1658 with Kriging and 0.9049 with BCSD as InternAgent-1.5 optimized a deep-learning downscaling method.The method better preserves fine-scale spatial gradients and localized precipitation maxima than interpolation approaches that smooth spatial fields.Source: §3.4.1 · p. 23, Table 9 · p. 23

3.4.2. Life Science

• InternAgent-1.5 narrowed 125 HCC candidate genes to GPR160, linking its tumor-specific overexpression to poor recurrence-free survival and potential immune-related signaling.It integrated differential expression, mutation and copy-number, survival, pathway, and tractability evidence from resources including GEPIA, TCGA, GEO, and OpenTargets.Source: §3.4.2 · p. 23
• ARG2 ranked first as a CRC target, with a proposed mechanism linking mitochondrial arginine depletion to reduced nitric oxide, impaired T-cell function, and tumor-promoting polyamine biosynthesis.The system generated dose-response, organoid-validation, and immune-profiling recommendations; ARG2 inhibition showed dose-dependent anti-tumor effects in HCT116 cells and multiple CRC PDO models.Source: §3.4.2 · p. 24

3.4.3. Biological Science

• Wet-lab assays and dry-lab predictions found sfGFP had high functional brightness and reliable folding efficiency, and InternAgent-1.5 produced a final report identifying variants for downstream use.The workflow included fluorescence assays, stability measurements, quality-control checks, and structural-stability and sequence-function predictions coordinated through SCP.Source: §3.4.3 · p. 25

3.4.4. Physical Science

• 0.86 Top-1 accuracy and 0.94 Fingerprint Tanimoto Similarity for major-product prediction; by-product Top-1 accuracy is 0.62, exceeding the cited reasoning-model comparisons [62].The system uses RDKit descriptors and atomic-conservation reasoning to infer products and by-products.Source: Table 10 · p. 26, §3.4.4 · p. 26
• For DprE1 inhibitor design, InternAgent-1.5 proposed piperidinopyrimidine bioisosteres, then replaced a lipophilic piperidine with morpholine and added para-fluorination to address solubility and metabolic stability.Candidate refinement used calculated Synthetic Accessibility, Tanimoto Similarity, and LogP metrics.Source: Figure 11 · p. 26, §3.4.4 · p. 27

Task-Episodic Memory

• With TEM active, performance trajectories rise smoothly, whereas removing it yields irregular or stagnant progress and repeated unproductive strategies.Retrieved episodes provide evidence from earlier trials, helping avoid ineffective choices and refine hypotheses; Figure 12 presents long-horizon optimization experiments illustrating stable iterative improvement.Source: §3.5 · p. 27, Figure 12 · p. 27

Semantic-Knowledge Memory

• With SKM enabled, research objectives reflect accumulated methodological knowledge, avoid saturated conceptual regions, and preserve semantic continuity and novelty across exploration batches.Figure 13 illustrates objective evolution from an initial seed, balancing exploration breadth with strategic redundancy that deepens promising sub-domains.Source: Figure 13 · p. 27, §3.5 · p. 28

Strategy-Procedural Memory

• The full system achieves higher success rates and more coherent plans than the SPM-ablated baseline, while procedural priors reduce unnecessary branching and redundant tool calls.Without SPM, plans become longer and fragmented, with error propagation and imprecise tool-call parameters.Source: §3.5 · p. 28, Table 11 · p. 28

4.1. Agentic AI for Scientific Discovery

• Scientific-discovery agents automate hypothesis and experiment design, evolve algorithms, coordinate specialized research agents, or unify retrieval, experimentation, and theory development,,,,.Robin likewise integrates planning, data analysis, and validation in a closed loop that discovers compound candidates without manual intervention.Source: §4.1 · p. 28

4.2. Deep Research Agents

• Deep Research agents progress from tool-integrated retrieval toward adaptive, iterative, multimodal long-horizon workflows, with unified single-agent loops and modular multi-agent designs [87], [88],,,,, [91], [68], [92], [70],, [93],, [94], [95].Recent studies also show the value of explicit structures and self-evolving mechanisms for multi-hop reasoning.Source: §4.2 · p. 29

4.3. Memory Mechanism

• Agent memory spans token-level, parametric, latent, short-term interaction, and long-term episodic approaches, but existing work chiefly targets shorter-horizon interaction settings, unlike scientific workflows [96],, [98], [99], [48].These memory techniques support incorporating prior information, while remaining orthogonal to multi-stage scientific discovery workflows.Source: §4.3 · p. 29

• InternAgent-1.5 unifies generation, verification, and evolution with deep research, solution refinement, and long-horizon memory for cross-disciplinary scientific workflows.Source: §5 · p. 29
• Evaluations show strong structured scientific reasoning, competitive algorithmic solutions, and multi-step computational and empirical workflows aligned with scientific principles and real-study findings.Source: §5 · p. 29
• Future work aims to strengthen computational reasoning–experimental validation coupling and accelerate the path from generated hypotheses to verifiable results.Source: §5 · p. 29

A.2. Earth Science example

• The InternAgent-1.5-generated report asks whether the AMOC will cross a 21st-century tipping point amid projections of gradual weakening and possible abrupt collapse.It frames the divergence around paleoclimate proxy evidence for early-warning signals and climate models whose biases may overstate AMOC stability.Source: §A.2 · pp. 36–37
• The AMOC transports about 1.25 petawatts of heat and supports carbon sequestration and nutrient distribution; disruption risks Northern Hemisphere cooling, rainfall shifts, and U.S. East Coast sea-level rise.The report links its heat transport to relatively mild Northern Hemisphere climates and its carbon and nutrient cycling to ocean and ecosystem functions.Source: §A.2 · p. 36
• The AMOC’s northward warm, salty upper flow and southward cold, dense deep flow depend on thermohaline density contrasts, freshwater inputs, and wind-driven transport and mixing.Freshwater can stratify surface waters and inhibit deep convection, while winds help drive horizontal currents and deep-water upwelling.Source: §A.2 · pp. 37–39
• The mainstream projection is high-confidence gradual AMOC weakening, whereas abrupt collapse before 2100 remains a low-probability, high-impact risk amid climate-model uncertainty.The report distinguishes the system’s abrupt shifts in paleoclimate from its projected anthropogenic industrial-era weakening.Source: §A.2 · p. 41
• Paleorecords contrast glacial abrupt AMOC shifts associated with freshwater-driven disruption of deep-water formation with comparatively subdued Holocene variability, although freshwater sensitivity persisted.The report describes freshwater stratification inhibiting deep convection during glacial slowdowns and notes a Holocene weakening event associated with the 8.2 ka climate event.Source: §A.2 · pp. 41–42

The Modern Era: Detecting Trends Amidst Natural Fluctuation

• Proxy studies infer 15–20% AMOC weakening since the mid-20th century, but reconstructions are disputed; direct monitoring began in 2004 and includes a temporary 30% decline in 2009–2010.The short direct record makes it difficult to distinguish a long-term forced trend from natural low-frequency variability.Source: §A.2 · p. 42

Projected 21st-Century Weakening and the Risk of Collapse

• CMIP6 projects a 34% to 45% AMOC decline by 2100 under high emissions, while IPCC assessments project 21st-century weakening with very high confidence.The projected decline corresponds to approximately 6 to 8 Sverdrups and is larger than projected by CMIP5; individual models vary.Source: §A.2 · p. 43
• The IPCC assesses abrupt AMOC collapse before 2100 as very unlikely but with only medium confidence, reflecting model limitations and evidence of modern instability.Models may overestimate stability and underestimate sensitivity to freshwater forcing; some statistical analyses instead suggest a tipping point within decades.Source: §A.2 · pp. 43, 44, 47

Statistical Early-Warning Signals in Observational and Paleoclimate Data

• Ditlevsen’s statistical analysis forecast collapse around 2057, with a 95% confidence interval of 2025–2095, after finding rising SST variance and autocorrelation since the 1970s.Figure 2 contrasts timing estimates and uncertainties from statistical early-warning analyses, specialized model experiments, and CMIP6 ensembles.Source: §A.2 · p. 48
• Shin et al. found rising SST and salinity autocorrelation in the eastern SPNA under current conditions, while paleoclimate records support autoregressive signals as early warnings [17].The observed signals included a 2014–2019 freshening event; the study argues that ESMs may overestimate AMOC stability.Source: §A.2 · pp. 48–49

Physically-Based Precursors from Specialized Models

• A CESM experiment identified southward freshwater transport at 34°S as a physical precursor; reanalysis extrapolation estimated mean collapse in 2050 and 59% (±17%) probability before then [18].The experiment induced collapse with sustained freshwater forcing over 1,700 years, judged unrealistically high but needed to counter the model’s stability bias.Source: §A.2 · p. 49

Methodological Debates and Countervailing Evidence

• Critics argue that short, potentially non-stationary time series and a late-1990s structural break undermine rolling-window early-warning methods; an alternative analysis found no impending-collapse signal.The alternative used a Hodrick–Prescott filter and cumulative plotting on the same reanalysis data.Source: §A.2 · p. 50
• Extended CMIP6 runs showed shutdown in all nine models, while Antarctic meltwater and Southern Ocean upwelling were reported as stabilizing feedbacks; a 34-model analysis deemed 21st-century collapse unlikely [14][20].These findings highlight a divide between tipping-point analyses and broader model ensembles that indicate substantial, though potentially limited, resilience.Source: §A.2 · p. 50

Challenges in Statistical Analyses of Paleoclimate Proxies

• Paleoclimate proxies are indirect, noisy, and geographically limited; single-core records have limited spatial and temporal resolution and may be dominated by local or regional variability, masking or mimicking basin-scale AMOC signals.In marine sediments, bioturbation acts as a low-pass filter that can artificially increase autocorrelation and generate a spurious EWS artifact rather than a true climate signal.Source: §A.2 · p. 51

Uncertainties in Data Interpretation and Reconstruction

• Uncertain age models can distort EWS trends, and proxy-calibration choices or unpropagated uncertainty can make reconstructions misleadingly precise [26].Red noise also has autocorrelation that can be difficult to distinguish from a genuine early-warning signal [17][27].Source: §A.2 · p. 52

The Role of Methodological Assumptions

• Stationarity assumptions and detrending choices can create phantom EWS or remove real signals, while collapse dates such as 2025–2095 depend on the assumed bifurcation model [10].The selected indicator also preconditions the outcome, helping explain why studies of the same data can reach opposing conclusions [10][19].Source: §A.2 · pp. 52–53

Intrinsic Stability Biases from Misrepresented Feedbacks

• Correcting freshwater-transport biases can make a previously stable AMOC model collapse under more realistic forcing, while the bias shifts its collapse threshold to an “unrealistic parameter regime” [29][1].Source: §A.2 · p. 53

Divergence from Incomplete Ocean-Ice Sheet Interactions

• Coupled Antarctic Ice Sheet melt can delay projected AMOC collapse by approximately 35 years, while Greenland melt destabilizes and West Antarctic melt stabilizes the circulation [20].The opposing effects make projections sensitive to the relative timing and magnitude of melt from each pole.Source: §A.2 · p. 54

Conclusion and Future Research Priorities

• ESMs project gradual, non-critical AMOC weakening, while proxy and observational analyses suggest a tipping point could be crossed this century.The report attributes this divergence to limitations in both modeling and observations and calls for better reconstructions and more physically realistic projections.Source: §A.2 · p. 57

Advancing Paleoclimate Reconstructions of AMOC Variability

• Reconstruction work should expand high-resolution archives in deep-water formation zones, validate direct circulation proxies, combine proxy types, and model post-depositional artifacts.The report notes that geographically sparse, coarse records and local variability can obscure basin-scale AMOC signals; traditional proxies can conflate temperature, ice volume, and salinity [24][10], while archive noise can generate spurious early-warning signals [23].Source: §A.2 · p. 60
• Age-model uncertainty, calibration choices, and unexplained variance warrant Bayesian chronology and probabilistic ensembles to distinguish faint early-warning signals from red noise.The report recommends integrated Bayesian chronological modeling and hierarchical Bayesian ensembles [27], rather than single best-fit reconstructions, to assess signal-to-noise [17].Source: §A.2 · p. 60
• Early-warning detection should resist structural breaks, use objective detrending and multiple AMOC proxies, and inform physics-constrained probabilistic risk assessments.The report recommends benchmarking detection algorithms against synthetic data and ESM output, and replacing simple bifurcation extrapolations with forecasts conditioned on deep uncertainties.Source: §A.2 · pp. 60–61

Refining Climate Model Projections of AMOC Stability

• Many ESMs simulate freshwater import into the Atlantic, contrary to observed export, reversing a destabilizing salt-advection feedback and making the AMOC appear overly resilient.The report recommends better atmospheric parameterizations, higher-resolution ocean models, and systematic bifurcation analyses to test whether modeled stability is genuine [29][38].Source: §A.2 · p. 61
• Climate models should replace idealized freshwater hosing with two-way Greenland and Antarctic ice-sheet coupling, since melt from the poles can have opposing effects on AMOC stability.The report notes that Antarctic melt can stabilize the AMOC and delay collapse by decades, and recommends fully coupled, bi-hemispheric cryosphere simulations [1][20].Source: §A.2 · pp. 61–62
• Projections should represent the full freshwater budget and route localized melt through high-resolution subglacial hydrology and plume models.Current models can omit geothermal basal melt and dilute meltwater by spreading it uniformly across large grid cells, obscuring its effects on stratification and local convection [31].Source: §A.2 · p. 62