Latest Scientific Breakthroughs of 2026, Scored and Ranked

The research published in 2026 so far, scored on evidence, impact and novelty, as of 11 November.

On 11 November 2026 Frontier's index held 368 published entries, and 295 of them are papers published this year. This report ranks those 295. Their Frontier Scores run from 38 to 78 out of 100, with a median of 63, and the highest belongs to a study of Germanic languages through ancient genomes, which sequenced 712 ancient genomes and suggests the branch may have reached Scandinavia some 800 years later than assumed, at 78. No entry comes near 100, by design: a perfect score would take a result that is at once established, field-changing and wholly new.

Two rules shape the ranking. It covers papers published in 2026: the 73 entries published earlier are in the index but not in this year's table, and the highest of them scores 52. And the Frontier Score ranks the active edge of science: it multiplies each result's weighted pillars by a recency factor, so a paper published in mid-September keeps about 93% of its standing today while one from mid-May keeps about 80%. The live top therefore skews toward the newest papers, which also make up most of the index's 2026 coverage (254 of the 295 were published from September on). Section 5 shows the same year with the recency factor taken out.

Every score here is the live index's on 11 November. The index re-scores each entry weekly as citations and other evidence arrive, so a live page can differ from the number printed here. This edition replaces the one of 27 October, when the index held 297 entries, 222 of them from 2026, and its top score was 75 (MemBrain v2, now 74); the 31 August edition before it had 101 entries and a top score of 70.

The 2026 ranking so far

Every paper published in 2026 that scores 70 or above, 29 in all, in the index's own order. Field and journal are the entry's; the last column says what the paper found, with the qualifier that matters, and links to the paper.

Biomedicine supplies 18 of the 29 and Nature publishes 15; 20 of the papers appeared in October and 5 in November, which is where the recency factor and the index's own intake both point. Every one of them is rated Early signal or Firming up, the confidence labels for work with little citation history yet: their Impact is an estimate, and their places will move as other researchers cite, test and build on them. The full ranking, every entry and every pillar, is live and sortable on the Frontier Index.

Contents

1. The shape of 2026 so far
2. How Frontier scores a breakthrough
3. The results leading 2026
4. Field by field
5. What the scores reveal
6. The outlook: what to watch into 2027
7. How to read the index yourself

Each numbered section stands on its own, so you can begin with the field that interests you and double back. The underlying table is downloadable in full from the open dataset, so nothing below has to be taken on faith, and where the data is uncertain we say so.

1. The shape of 2026 so far

An evidence score holds apart the reasons a result matters, attaches each to public signals anyone can retrieve, and only then combines them, so the final number carries its own reasoning. The most recent editorial lists are, by their authors' own admission, best guesses rather than measurements. One prominent technology list calls its selections "educated guesses" and notes candidly that it does not always get them right. - MIT Technology Review's 10 Breakthrough Technologies 2026 There is nothing wrong with an honest guess from expert editors. The difference is that a score built from public records can be audited and re-run when new evidence lands.

Start with where the 2026 entries come from in time, because it shapes everything else. The index went public at the end of August with 101 entries, many of them from 2025. Since then it has taken in new papers as they appear, daily since early October and up to six a day since 7 November: mostly papers from the last two weeks of the journals it reads, with a smaller pass for papers from roughly the past year that are already being cited. The picks are made by score, which favours fresh work, so the earlier months fill in slowly. Its 2026 coverage is therefore concentrated: 254 of the 295 entries were published from September on, and only 41 in the eight months before.

That concentration is the first thing to keep in mind when reading this as a report on the year. A September or October paper had a good chance of entering the index; a paper from March entered only if it was in the launch set or later stood out among the past year's already-cited papers. The ranking is complete for what the index holds, not for everything 2026 published, and the most important results of the first half of the year are represented by a few dozen entries, among them the AI co-scientist ( paper) that led the August edition, Evo 2 ( paper) and the single base-editing infusion that cut LDL cholesterol ( paper).

Within the 295, the scores are compressed near the top. The mean is 62.2 and the median 63; 219 score 60 or more, 107 score 65 or more and only 20 fall below 50. Two features of the method explain the shape. New entries are published automatically only when they first score 40 or more, which keeps the bottom of the scale nearly empty. And the recency factor is still close to 1 for papers a few weeks old, so the newest entries keep nearly all of their standing while the year's older papers have already lost some: entries from September on average 63.6, those from January to August 53.1.

The index as a whole, 368 entries including the 73 published before 2026, averages 58.0 with a median of 61. The August edition reported a mean of 45.7 for its 101 entries. The difference is mostly composition: the index now holds far more fresh papers, whose recency factor is near 1, while the entries published before 2026 average 41.1.

Biomedicine is the center of gravity. With 112 of the 295 2026 entries, Biomedicine and Health holds more than a third of the year, followed by Neuroscience with 46, Climate and Environment with 31, AI and Computing with 26 and Materials and Energy with 25. This largely reflects how much biomedical research the selective journals the index reads publish, not a judgment that biology matters more than physics.

2. How Frontier scores a breakthrough

The hard problem behind any index like this is not opinion, it is commensurability. A clinical trial, a nuclear clock and a toolkit for reading genomic models are not obviously measurable on the same axis, and the easy move is to rank them by feel. Frontier's answer is to route every candidate through the same fixed pipeline, so that a neuroscience result and an astronomy result face identical questions, drawn from identical kinds of public evidence, before either receives a number. The score is the output of a repeatable process, and that process is the same for all 368 entries, which is what makes two very different results comparable at all.

That process rests on three pillars, each answering a separate question about a result. Evidence asks whether the finding is established and verifiable. Impact asks whether it matters, judged against its own field and by how fast it is being taken up now. Novelty asks whether the contribution is genuinely new. They are held apart because a result can be strong on one and weak on another, and collapsing them too early is exactly the error an evidence score exists to prevent. The pillars are computed from 13 independent public data sources. New papers enter every day, up to six a day, and every published score is re-checked against its sources each week, so a score shifts as the evidence around a result accumulates rather than freezing at publication. The path from raw literature to a ranked entry runs in five stages.

Each stage is mechanical and inspectable. Discovery reads the newest papers from leading journals and preprint servers in OpenAlex every day and takes in up to six, first the papers their own journal published a commentary on, then the strongest of the rest, rotating across fields; an editor can also add a paper the sweep missed. Enrichment gathers the signals attached to each work from the other sources: citation records in four more indexes, venue, openness and integrity records, reach beyond citations, and the research topics it combines. Scoring turns those signals into the three pillar values. Ranking combines the pillars into a single 0 to 100 Frontier Score. Publishing exposes the entry with its figures linked to their primary records. The inputs are the ordinary infrastructure of open scientometrics, and they are named on our sources page.

• OpenAlex, the fully open index of the global research literature, built by a nonprofit as the successor to Microsoft Academic Graph.
• Crossref and DOI records, anchoring each result to a citable, permanent identifier, and recording corrections and retractions.
• NIH iCite, supplying field-normalized citation measures for the biomedical literature.
• OpenAlex topic co-occurrence counts, used to judge how atypical a work's combination of research topics really is.

Naming the sources is what makes a score reproducible rather than merely reported. The open catalog at the base of the pipeline is itself a public, documented system anyone can query. - the OpenAlex paper Anyone who disagrees with where a result landed can pull the same records, recompute the same pillars and find the exact place their judgment departs from Frontier's. The source register and the current methodology, version 1.3.1, are public, including every change since version 1.3.0 and what it moved.

Evidence rewards peer review and the selectivity of the venue (a fresh paper in one of the five general flagships, Nature, Science, Cell, NEJM and the Lancet, reads stronger than one in a selective specialist journal, which reads stronger than one in a high-volume mega-journal), the breadth of independent records that hold the work, openness, and a clean integrity record. Because the index draws on selective journals, Evidence runs uniformly high: the 2026 entries average 90.5 on it, and it does less than the other two pillars to separate the strong entries from the merely good ones.

Impact is where the measurement gets careful, because raw citation counts punish slower-citing fields and flatter faster ones. Frontier anchors it on field-weighted measures that compare a result with the world average for its own field, year and type, alongside how fast the work is being cited right now and how many citations genuinely build on it. Lifetime citation volume is deliberately not rewarded. - the field-weighted citation impact definition A second, article-level lens, the NIH Relative Citation Ratio, benchmarks a biomedical paper against its own co-citation neighborhood rather than a broad disciplinary average. - the Relative Citation Ratio in PLOS Biology

Novelty builds on a well-known finding that the highest-impact science tends to pair conventional ideas with atypical ones, which Frontier measures through how rarely a work's research topics have been paired before. - the atypical-combinations study in Science Recency, rising attention and being at the edge of peer review add to it. A result that fuses two rarely joined literatures reads as more novel than one that extends a single crowded line of work.

The pillars then combine with fixed weights: Novelty carries half of the score and Evidence and Impact a quarter each. The weighted result is multiplied by a recency factor, because Frontier ranks the active edge of science: a result keeps 75% of its standing at eight months old and about 63% at sixteen, and never falls below half, so a settled landmark is retained but no longer leads. A retraction sets the score to zero. Every entry page shows this arithmetic, the weighted pillars and the factor applied to them.

One more layer changes what a reader sees but not the ranking. Each entry carries a separate confidence reading, built from how many citations it has gathered across the indexes, how long it has been out, how closely the indexes agree and how many independent kinds of evidence back it, among other things: Settled, Firming up or Early signal. The tier label is capped by it. An Early signal entry is labelled at most Notable, and a Firming up entry at most Major, whatever its score, so a weeks-old paper cannot wear a Landmark badge before anyone has had time to cite it. That is why 275 of the 295 entries published in 2026 carry the Notable label, although 219 of them score 60 or more, the Landmark line, and not one is labelled Landmark.

3. The results leading 2026

Twenty-nine papers published in 2026 score 70 or above. The twelve highest are profiled here, by theme rather than strictly in rank order, with each one's score, the pillars that explain it, what it found and what it does not show. Every one is a peer-reviewed journal paper, eleven of the twelve in Nature or one of its sister journals, and every one but MemBrain v2 is an Early signal: little citation history yet, so its Impact is an estimate that will move. Six of the twelve carry the same publication date, 28 October, and sixteen of the twenty-nine at 70 or above were published on that day or later.

Two origin stories

The highest Frontier Score of 2026 so far, 78, belongs to an ancient-genomics study: tracing the spread of Germanic languages, which sequenced 712 ancient human genomes from the Late Neolithic to the Viking Age and analysed them alongside 4,009 previously published ones. They reveal a population that is hard to see in the archaeological record, entering Sweden likely from the Baltic region by around 4000 years before present and widespread throughout Scandinavia by 3500. By 1650 years before present the same ancestry moves south into presumed Celtic-speaking areas, including present-day Germany, Poland and the Netherlands; during the Migration Period it turns up among Germanic-speaking Anglo-Saxons in Britain and Langobards in southern Europe, and a large-scale back-migration north into Denmark and southern Sweden coincides with the linguistically documented emergence of Old Norse. Germanic's arrival in Scandinavia is usually tied to the Corded Ware cultures, so this would place it some 800 years later than traditionally assumed. Its pillars are Evidence 93, Impact 59 and Novelty 82.

What it shows is genetic ancestry, not language. The abstract calls the Baltic-linked group a potential vector, and the authors write that they cannot rule out more complex scenarios, for instance Palaeo-Germanic spreading by cultural diffusion with no genetic impact, in the opposite direction. The group's origin is inferred rather than sampled: the authors place its formation somewhere between Finland and northeast Sweden, a region and period from which no genomes yet exist, and only 578 of the 712 new genomes survived filtering into the final imputed dataset. Paper

A far older origin story scores 74 (Evidence 92, Impact 60, Novelty 74). Eukaryotes' closest relatives reports the isolation of two archaeal strains, HC1 and SC1, from Heimdallarchaeia, the lineage that branches closest to eukaryotes. Both are obligately anaerobic, peptide-degrading and syntrophic, grown with methanogenic partners, and both produce protrusions and vesicles but lack persistent internal membrane structures. They grow extraordinarily slowly: HC1 doubles in about 7 to 12 days, SC1 in about 14 to 20. HC1 carries genes associated with aerobic life, including cytochrome c oxidase, yet its co-cultures cannot grow under oxic or microoxic conditions alone, which the authors read as the oxidase detoxifying oxygen rather than respiring it. Because those features are widely conserved across the phylum, they argue the archaeal ancestor of eukaryotes may have shared them. That last step is an inference from conservation, not a fossil or an ancestral reconstruction, and the oxygen conclusion rests on what the cultures do rather than on measuring the enzyme. Paper

A toxin, a receptor and an antibody

At 77, and with the highest Novelty of any 2026 entry, 91, is a toxin behind necrotizing enterocolitis, the most devastating gastrointestinal emergency in preterm infants and a cause of lasting neurodevelopmental impairment. The authors find Clostridium perfringens pathobionts carrying Delta toxin in the gut microbiome of infants with severe and fatal disease; serological testing of severe surgical cases found 7 of 12 positive for antibodies to the toxin, and no infant without the disease tested positive. In neonatal mice, infection with a Delta-toxin-positive clinical isolate, or purified toxin alone, reproduced both intestinal necrosis and acute cerebral injury. The toxin preferentially targets macrophages, activating inflammasome signalling and pyroptosis. A genome-wide CRISPR screen identified the tetraspanin CD9 as its host receptor: knocking CD9 out of human cells cut toxicity about 40,000-fold, deleting it in mouse myeloid cells protected the animals, and a neutralizing antibody protected against disease progression.

The human half of that case is association in a small cohort. The Delta-toxin-positive isolates came from two deceased patients, the outcome comparison rests on two cases against five, and the causal chain is mouse work. The authors state the limits themselves: no isogenic Delta-toxin-deletion strain to confirm the pathology is toxin-dependent, and multicentre cohorts needed before anyone knows how common these strains are. The brain injury also looks indirect, since toxin was barely detectable in the brain and the authors read the damage as driven by systemic inflammation. Its Impact of 37 is the lowest of the twelve, and below the median of 51 for the other sixteen entries published the same day, because Impact this early is an estimate rather than a tally. Paper

Load, fibril by fibril, inside an intact disc

The only physics entry in the top twelve is second overall at 77 (Evidence 91, Impact 47, Novelty 85). Multimodal X-ray imaging of hierarchical fibre mechanics introduces TomoSAXS, a full-field three-dimensional small-angle X-ray scattering tomography which, combined with phase-contrast tomography, maps collagen mechanics from the fibril to the fibre scale across a whole tissue. On 3 intact rat spinal discs imaged before and after a 50 micrometre compression, it reconstructed nanoscale structure for more than 55% of all fibres in the scanned volume, and found that the molecular pre-strain inside fibrils is patterned by the disc's layers and correlates with fibre strain: fibrils under initial tension tended to compress under load, those under initial compression to stretch, which the authors read as load-sharing through molecular unwinding. One whole human disc was imaged too, without loading.

It is a methods paper resting on few samples, with no randomization or blinding and no adjustment for multiple comparisons. The regressions behind the load-sharing picture are modest and come from a single representative disc, strains were recorded after the tissue had relaxed, and the human disc was formalin-fixed, which will have affected its nanoscale collagen. The suggestion that fibre curvature may partly explain why the posterior regions of the disc herniate most comes from rat discs and is not a clinical finding. Paper

Four maps of what genetic variation does, cell type by cell type

The densest cluster in this year's top is one question asked four ways: which variants change gene activity, and in which cells.

A population-scale muscle atlas scores 75 (Evidence 93, Impact 57, Novelty 77). Profiling chromatin accessibility and gene expression in single nuclei from 287 human skeletal muscle biopsies, it identifies 13 cell types, more than 7,000 expression QTLs and more than 100,000 chromatin accessibility QTLs, and colocalizes almost 3,500 GWAS signals across 38 traits with them, most in a cell-specific way. Its causal directional maps are statistical inferences from genetic variation rather than experiments, and the authors write that orthogonal testing to confirm an impact on disease remains a critical step. Paper

Two Nature papers of 28 October, each drawing on 1,925 individuals, come at it from different angles. Tandem repeat variation, at 74 with an Impact of 67, pairs whole-genome and single-cell RNA sequencing across more than 5.4 million blood cells to show that repeated stretches of DNA, usually skipped by studies that read single letters, regulate immune genes cell type by cell type: more than 69,000 associations, 30.7% specific to one of 28 cell types, with fine-mapping flagging 1,490 repeats as candidate causal drivers. Paper Its companion, phase 1 of TenK10K, also at 74 and with an Evidence of 94, finds putative common regulatory variants for 83% of all 21,404 genes tested and cumulative rare-variant signals for 47%, with about half of the effects appearing in only one or a few cell types. Paper Both are statistical associations in blood from cohorts of European ancestry, and both sets of authors say laboratory validation and more ancestry diversity are what comes next.

The fourth is the machinery. SAIGE-QTL, at 74 (Evidence 90, Impact 53, Novelty 78), is a Poisson mixed model that works on single-cell read counts directly instead of averaging each donor's cells into a pseudobulk. In 1.2 million immune cells it found 48.8% more genes with a significant eQTL than the pseudobulk pipeline it was compared with. It is not faster than pooled analysis, two pseudobulk tools ran about four times quicker, and more detected eGenes means more detected associations, not identified causal variants. Paper

Software that reads biology, and papers that answer back

MedGemma, at 74, carries the highest Impact of any entry at 70 or above, 94, against a Novelty of 64, the lowest of the twelve. It is a collection of open medical vision-language models built on Gemma 3 at 4 and 27 billion parameters, improving on their base by 15.5 to 18.1% at chest X-ray finding classification and scoring 86.2 on MedQA. These are benchmark results, which the authors call only a first step: some benchmarks are near saturation, older public ones may have leaked into training, and the chest X-ray report ratings come from a single unblinded radiologist, who judged 49% of reports on abnormal studies equal to or better than the original, meaning the rest were worse. Paper

MemBrain v2, the top score of the previous edition, is now 74 with an Impact of 93 and remains the only entry at 70 or above whose confidence has passed Early signal, to Firming up. It unifies membrane segmentation, particle localization and quantitative analysis in cryo-electron tomography into one pipeline. As a method paper its value depends on how well it generalizes beyond the datasets it was shown on, and on adoption. Paper

Paper2Agent, at 73, is the only one of the previous edition's leaders whose score went up. It converts a paper and its codebase into an AI agent acting as a virtual corresponding author, exposing the paper's data, code and workflows as tools a chat assistant can call; in case studies on AlphaGenome, Scanpy and TISSUE the agents reproduced the original results and handled new queries. Those are computational papers with code available, so how the approach fares without runnable code is open. Paper

What curiosity looks like in children and in agents

At 73, intrinsically motivated exploration in an open world put adults, children aged 8 to 12 and AI agents in the same open-ended game and measured three objectives AI researchers use to make agents curious: entropy, information gain and empowerment. All three correlated positively with how far people got. Across agents and humans, entropy rose quickly and then plateaued while empowerment kept climbing, which the authors read as early exploration seeking novel states and later exploration seeking states the explorer can control. These are correlations in a small sample, 24 adults and 26 children, not a test of what causes exploration. Paper

How the leaders earn their scores, and how fast they turn over

Set side by side, the pattern is plain. Evidence is high and nearly level: across all twelve leaders it runs from 88 to 94. The recency factor barely separates them either, from 0.918 for the oldest to 0.994 for the newest. What separates them is Novelty, which carries half the score and spans 64 to 91, and Impact, which spans 37 to 94 and is still an estimate for almost all of them. MedGemma and MemBrain v2 are the exceptions that make the rule: both lead on Impact with Novelty near the bottom of the twelve, and both sit at 74 rather than at the top.

The other thing the pillar data says is that this list turns over fast. Of the 18 papers profiled in the edition of 27 October, 11 now score below 70: the Honduran village-network trial (74 to 69), the designed metalloproteases (72 to 69), Angelman syndrome stress signalling (71 to 68), the cold-activated BH2 network (71 to 68), Tangermeme (71 to 69), cohesin and euchromatic domains (71 to 69), the DNA Typewriter lineage record (70 to 69), retinal prediction of hypertensive disorders (70 to 67), the lifespan prefrontal atlas (70 to 68), agricultural liming as a carbon sink (70 to 69) and the State virtual-cell model (70 to 69). Only two of the eighteen are still in the top twelve.

Nothing collapsed. The largest fall in fifteen days is five points, and most moved by one or two. Two forces do it. The recency factor erodes every week, so a paper has to gain on a pillar simply to stand still: the eye-movement study still holds Novelty 88, the second highest of any 2026 entry, and has slipped from 72 to 70 anyway. And Impact at this age is a field-weighted estimate rather than a tally, so it is the least settled pillar in every one of these scores: the village trial now reads Impact 29 against an Evidence of 96, while the State virtual-cell model reads Impact 73, higher than ten of the twelve current leaders, and still sits at 69 because its Novelty is 67. Paper2Agent is the only one of the eighteen to gain enough on its pillars to outrun the erosion. With 29 papers packed between 70 and 78, a point or two reshuffles a great many places.

Strip the recency factor out and a different year appears. The highest weighted score of any 2026 entry is 82, for the AI co-scientist system published on 19 May, whose recency factor of 0.808 leaves it at 66 today, outside this section entirely. Among the leaders here, MemBrain v2's weighted 80.8 is the highest, ahead of the Germanic genomes at 79. Section 5 takes that view of the year in full.

4. Field by field

The index now covers ten fields rather than nine, because Mathematics has its first 2026 entry. The other change since the last edition is that the fields have converged. Averaged over their 2026 entries, the fields holding five or more of them sit within 6.5 points of each other, from 64.0 at the top to 57.5 at the bottom. Every one of them clears the same high Evidence bar, because the index draws on selective journals and traces each entry to a primary record: their Evidence averages run from 87.8 to 92.3. So Evidence cannot be what separates one field from another. What does is Novelty, which carries half the score, and Impact, which for most 2026 papers is still an early estimate. Three groups are too small to average honestly: Mathematics holds one entry, Other frontiers seven and Quantum eight, so a single paper moves each of those figures, and a field of one is a paper's score rather than a field's standing. Read the chart as a map of where new and consequential work concentrated this year, not as a scoreboard of which sciences are healthier.

Biomedicine and Health

The largest field by far, with 112 entries published in 2026, an average of 63.5 and a Novelty average of 66.6. Its ceiling is the spread of Germanic languages traced through ancient genomics at 78, whose genomes point to a later group, likely from the Baltic, as a potential vector, placing Germanic's arrival in Scandinavia some 800 years later than traditionally assumed. The entry is careful that this is genetic ancestry and not language, and that the authors cannot rule out more complex scenarios. Behind it, at 77, a Clostridium perfringens toxin that targets CD9 appears to drive a subset of the deadliest gut emergency in premature babies, and the brain injury that follows it: antibodies turned up in 7 of 12 infants with severe disease and in none without it, and purified toxin alone reproduced both injuries in neonatal mice. The human evidence there is association in a small cohort; the causal chain is mouse work. At 75, population-scale single-nucleus profiling of skeletal muscle matched almost 3,500 trait signals to regulatory variants across 287 biopsies, most of them to variants that change how open DNA is rather than gene expression, by statistical inference rather than experiment. The year's clinical trials again score lower, chiefly on Novelty: the 19 entries from the New England Journal of Medicine, the Lancet journals and JAMA average 60.8 (Novelty 63.1), against 64.0 (Novelty 67.4) for the other 93 entries in the field. The highest of them is intermittent or continuous renal-replacement therapy at 68, followed by an AI assessment of national cancer control plans at 66, a global analysis rather than a trial, and screening with a multicancer early-detection test at 65. Biomedicine holds 18 of the 29 entries scoring 70 or more, and six of the ten fields have none at all.

• The Germanic genomics study in Nature
• The Delta toxin study in Nature
• The skeletal muscle profiling study in Nature Genetics

Neuroscience

46 entries, averaging 63.5, with the highest Evidence average of any field, 92.3, and a Novelty average of 67.0. Two papers tie at its ceiling of 71. Connectome simulations of fly walking narrow the circuit that makes stepping rhythms to three interneurons, with two of the simulation's predictions holding up in living flies; the circuit is a simulation result, which is why the authors call it putative. The functional architecture of the human insula, mapped in 87 people with brain electrodes, finds four regions, the front one influencing the rest, on evidence the open preprint itself calls a proxy measure of causal interactions. Behind them, at 70, mice that only watched fights later fought more like experienced fighters and shared their brain changes, though the decoupling at the centre of it is a statistical result rather than a measured synapse, and a population code for word meaning in the human hippocampus aligned most closely with GPT-2 embeddings. Neuroscience is the field whose Evidence is strongest and whose claims most often rest on inference from a model.

• The fly connectome study in Nature
• The human insula study in Nature
• The aggression observation study in Nature

Climate and Environment

31 entries, averaging 59.4, with an Impact average of 37.8, among the lowest in the index, and no entry reaching 70. Two share its ceiling at 69. The anthropogenic fingerprint on emerging infectious diseases tests many diseases and drivers together and finds emergence to be multi-causal, with the effects of deforestation, warming and agricultural intensification varying widely between diseases; its most consistent pattern is about what gets recorded, with reported outbreaks declining by a median of 32% for each additional hour of travel from a health facility. Agricultural liming in the Mississippi River Basin comes out as a net carbon sink on decade-to-century timescales once the counterfactual is the acidity inputs rather than the lime, a result for one basin that depends on which counterfactual is used. At 66, Caribbean anoles of closed forest raised their heat tolerance by 2 °C on average in a decade while open-habitat species barely changed. Climate results tend to be single basins, single countries or single decades, and their entries are explicit about it.

• The infectious disease drivers study in Nature
• The agricultural liming study in Nature
• The Caribbean anoles study in PNAS

AI and Computing

26 entries, averaging 64.0, the highest of any field with more than a handful of entries, and carrying by far the highest Impact average, 54.3. Its ceiling is MedGemma at 74, an openly released medical vision-language model in 4-billion and 27-billion-parameter sizes that improves on its base model by 15.5 to 18.1% at chest X-ray finding classification, with physicians rating 15.4% of its open-ended answers factually inaccurate against 18.5% for base Gemma 3. The authors call it a first step and nothing in it is clinical validation. At 73, research papers rebuilt as interactive agents turn a paper into a tested, tool-invoking agent a chat assistant can query, on evidence drawn from case studies of computational papers whose code and data are available. At 71, a language-model system doing sequential diagnosis diagnosed 83.2% of NEJM case-conference cases correctly against 36.1% for generalist physicians, in a controlled benchmark and not a trial in patient care. That pattern holds across the field: most of its evidence is benchmarks and the authors' own case studies, so AI and Computing is where independent evaluation matters most.

• The MedGemma study in Nature Medicine
• The paper-as-agent study in Nature
• The sequential diagnosis study in Science

Materials and Energy

25 entries, averaging 57.5, the lowest of the ten fields, with the lowest Evidence average, 87.8, and an Impact of 36.9. Two entries tie at its ceiling of 67. Atomically resolved exciton emission in MoS2 ties a defect glow often attributed to sulfur vacancies to carbon-hydrogen complexes instead, confirmed by cleaving off the hydrogen with a voltage pulse, in measurements made below 5 kelvin in ultrahigh vacuum. Free-standing two-dimensional metal oxides are made one atom thick by stripping the halogen out of layered oxyhalides with an electron beam, yielding single-crystalline CrO domains over 10,000 nm²; the predicted magnetic order below 230 K is density functional theory, not a measurement. At 66, twist-angle phase diagrams in moire bilayer MoTe2 link fractional topology, symmetry breaking, magnetic order and superconductivity to one knob across fifteen devices, while leaving the origin of that superconductivity open. Materials results are mostly mechanisms and methods whose applications lie ahead.

• The MoS2 defect study in Nature Nanotechnology
• The two-dimensional metal oxides study in Nature Materials
• The moire MoTe2 study in Nature

Physics

22 entries, averaging 62.1, with an Impact of 45.9 and a Novelty of 62.0. One entry alone reaches 70, and it sets the ceiling high: multimodal X-ray imaging of hierarchical fibre mechanics at 77, which measures collagen strain at the fibril and the fibre scale in one intact disc under load, with fibre strain rising with fibre curvature. It is a methods paper resting on three rat discs and one human one. Behind it, the first science results from ALPS II at 69 saw no axions and set a laboratory-only limit more than 20 times tighter than similar experiments, which is a bound rather than a measurement. At 68, polarized neutrons in MnF2 recast the archetypal antiferromagnet as an altermagnet, on a chiral signal the arXiv version puts at 5 to 11% of the scattering intensity. Physics in 2026 is precise measurement turning into new probes.

• The hierarchical fibre mechanics study in Nature
• The ALPS II results in Physical Review Letters
• The MnF2 altermagnetism study in Science

Space and Astronomy

17 entries, averaging 59.4, with Evidence at 91.0 and Impact at 41.4. Its ceiling is a Martian meteorite that formed at least 4.1 billion years ago at 67, a sample of early crust from a mantle source the collection had missed, whose altered minerals and oxygen isotopes also record water. One rock is one rock, as the entry says, and the age is a minimum. At 64, a gas-enshrouded black hole at cosmic dawn appears wrapped in dense gas 660 million years after the Big Bang, which would mean masses of such sources are overestimated, on modelling the authors themselves call highly simplistic. At 63, a double debris disc around a white dwarf is two rings where all 21 previously known such discs are one, attributed to two surviving planetary bodies that are proposed rather than observed. Space results are often single objects and interpretations awaiting confirmation, which keeps their Impact modest.

• The Martian meteorite study in Science
• The cosmic dawn black hole study in Nature
• The double debris disc study in Nature Astronomy

Quantum

Only 8 of the 2026 entries are quantum, averaging 60.1, so one paper moves the field's average. It carries the lowest Novelty average of the ten fields, 59.1. The top entry, at 67, is prethermal dynamical phase transitions and resonant melting, in which an atom simulator running up to 180 qubits maps long-lived dynamical phases and the resonances that destroy them; those states are long-lived rather than permanent, and the system is expected to thermalize beyond the timescales the experiment reaches. At 64, simultaneous single-qubit gates in a silicon spin-qubit array held fidelities above 99.99% with up to 3 qubits driven at once and above 99.9% with all 5, which are single-qubit gates on a 5-qubit array and far from a fault-tolerant processor. Helios, the 98-qubit trapped-ion machine that led quantum in August, now scores 58.

• The prethermal phases study in Nature
• The spin-qubit gates study in Nature Physics
• The Helios study in Nature

Other frontiers

7 entries that fit none of the named fields, averaging 62.9 on a sample small enough that one paper moves it, with the highest Novelty average in the index, 69.9, and the lowest Impact, 34.0. Its ceiling is an educational intervention that rewired village social networks at 69, a health-education trial across 110 Honduran villages that changed who people were tied to, in a pattern still visible 3 years after it ended, from one 22-month intervention in one country. At 65, two decades of American adolescents' outdoor time show boys' outdoor engagement falling sharply, erasing their long-standing lead over girls, in time diaries from 14,235 people aged 15 to 19. At 61, ancient proteins identify Denisovan remains in Southwest China, filling a geographic gap, though the postcranial evidence is a few fragments.

• The village networks trial in Nature
• The adolescent outdoor time study in PNAS
• The Denisovan proteins study in Nature

Mathematics

The index's newest field, and so far a field of one. The ABCD hypothesis scores 66, with Novelty 69: it proposes an answer to every comparison between infinite cardinal exponentiations, where the Continuum Hypothesis settles a single case, and argues the answer holds below Theta under Determinacy in L(R). Its reach is exactly its hypotheses, and the article describes what it contains as a sketch of the proof, with the stronger version left as a conjecture for forthcoming work. One entry is not an average, so Mathematics has a score on the chart and no standing yet.

• The ABCD hypothesis in PNAS

Put the ten back together and the ranking reads as a map of pillars rather than a league table. AI and computing gains from early citations and carries the index's highest Impact; the smallest groups and neuroscience lead on Novelty; materials sits at the bottom with both the lowest Evidence and nearly the lowest Impact; and fields whose results are single basins or single objects, like climate and space, sit low while their Evidence is close to the rest. The per-field breakdown, every entry and every pillar, is in the open dataset.

5. What the scores reveal

Once every entry is broken into three independent pillars, the index stops being a leaderboard and becomes a diagnosis. The single most useful pattern in the data is that the pillars behave differently, and that they trade places as papers age.

Across the 295 entries published in 2026, Evidence averages 90.5, Impact 44.9 and Novelty 65.0. Across the 73 earlier entries, the order flips: Evidence 91.3, Impact 75.7, Novelty 48.0. Evidence barely moves with age, because it is largely the venue bar and the records that hold a paper, and the index draws on selective journals. Impact grows as citations arrive, which takes months to years. Novelty falls, because part of it is recency and being at the edge of peer review. A 2026 paper is new and not yet cited; an older one is cited and no longer new.

That is why the live top of 2026 is held by papers weeks old, and why the same year looks different with the recency factor taken out. Each entry page shows its weighted pillar score before the factor and the factor applied to it. Ranked on the weighted pillars alone, the leader of 2026 is still the AI co-scientist of May, at 82.0, ahead of MemBrain v2 (80.8) and the Germanic genomes paper (79.0); the co-scientist's Frontier Score of 66 is that weighted score times a recency factor of 0.81. The single base-editing infusion for LDL cholesterol, at 77.6 weighted and a Frontier Score of 63, and Evo 2, at 76.7 and 57, are the clearest cases of strong results the factor pushes down the live ranking.

Neither view is the "right" one. The Frontier Score answers what is at the active edge now; the weighted pillars answer how strong a result's evidence, impact and novelty are regardless of when it appeared. A reader looking for the year's most consequential results should look at both, and at confidence.

The August edition's ten leaders show the recency factor at work. Every one has lost points since 31 August, by between 2 and 10. For most, the weighted pillars implied by their August scores are within a couple of points of today's, so the fall is the factor rather than a change in the evidence; the two exceptions move in opposite directions. The cortical output channel's weighted pillars are 4.5 points lower than its August score implied, and Helios's are 3.7 higher, the latter because methodology 1.3.1 stopped counting a publisher's own correction as an integrity flag. Since 27 October most moved by a point or none; the proton-coupled electron transfer study gained 3 and Evo 2 one, as citations reached their records.

Confidence tells a parallel story. Of the 295 entries published in 2026, 261 are Early signal, 26 Firming up and only 8 Settled. Among the 73 earlier entries, 61 are Settled. Freshness and certainty pull in opposite directions, and the score refuses to hide it: the papers leading this year's table are the ones least tested by the rest of science so far. That is not a flaw to correct but a fact to read with every score.

The same records explain part of the ranking. Several leaders carry a citation history from before their journal version, because OpenAlex merges each paper with its preprint: MedGemma, MemBrain v2 and Paper2Agent are the clearest cases, and each entry page shows the merged record its Impact is read from. That is a fair account of how a method spreads, preprint first, but it means those leaders are not new in the way the rest of the table is.

An honest index has to be pressure-tested against itself, so here are the places this reading is most open to challenge.

• Coverage is uneven. The index holds 41 entries from January to August against 254 from September on, so the first eight months are thinly represented, and a strong paper from March may simply not be in the index.
• The high Evidence floor is a selection effect. Evidence sits near 90 because a selective venue is the price of entry. "Evidence is rarely the bottleneck" is true inside the index, not a claim about research everywhere.
• Novelty may be undercounted. The novelty pillar rewards atypical combinations of research topics. A result that is genuinely new but built from conventional methods can still score modestly.
• Impact for a weeks-old paper is an estimate. Most 2026 entries have a handful of citations; their Impact will be re-read every week as the record grows.
• The ceiling is a calibration choice. A top score of 78 reflects a scale on which a result needs high marks on all three pillars at once, and a paper a few weeks old cannot yet have the citations a high Impact requires. Read it as a conservative method, not as a verdict that no 2026 result was field-changing.

None of these overturn the pattern; they sharpen how to read it. What remains scarce, on any reasonable measure, is work that is solid, new and consequential together, and that is the signal a ranked list of admired papers cannot give, because such a list never separates why a result earned its standing from the fact that someone decided it had.

6. The outlook: what to watch into 2027

A score is a snapshot, and the honest question a snapshot invites is where each line is heading. Three arcs run through the 2026 entries, and in each the evidence score describes the near term more modestly than the headlines do.

The first arc is AI inside the workflow of science. The 2026 entries include a framework that turns a paper and its code into an agent that reproduced the original results in case studies ( Paper2Agent, 71), a toolkit for reading the regulatory rules a genomic model has learned ( tangermeme, 71), a model that predicts how cells respond to perturbation ( State, 70), retinal-image prediction of hypertensive disorders in pregnancy ( Visionary AI, 70) and the AI co-scientist that led the August edition (66). What they share is that their evidence so far is largely the authors' own case studies, benchmarks and validation cohorts. The measure to watch is independent use: other groups reproducing results with these tools, and predictions confirmed in the laboratory or the clinic. The arc also runs the other way, with AI labs paying researchers who hold doctorates by the hour to write hard problems in their field and check what models answer, and the PhD roles listed on AITraining.jobs, a jobs index we also maintain, show who is hiring and at what pay.

The second arc is protein design reaching harder chemistry. The two designed proteases of October cut amide bonds, among the harder reactions to design an enzyme for, and their abstracts leave plain how far they are from use: most designs did not work, the best cysteine proteases still trail natural ones, and neither abstract gives selectivity figures against other proteins or reports any test in an animal. The next evidence that would move these entries is those figures, behaviour in animals, and the first applications.

The third arc is the brain mapped in fine detail. The 2026 entries include a lifespan atlas of the human prefrontal cortex, a single-cell atlas of vulnerability across brain disorders, a lipid atlas of the mouse brain and recordings of how hippocampal neurons encode word meaning. These are reference maps and measurements, and their impact will show in the studies that use them. For all three arcs the same reading applies: almost every entry is still an Early signal, and the scores will firm up, or fall, as other researchers cite, test and build on the work. That movement is visible week by week on each entry's page.

7. How to read the index yourself

The point of an evidence-scored index is that you do not have to take its word for anything. Every score traces back to primary records, and the index behind it is open: you can download it, inspect the method and check any number yourself. That is the difference between a ranking you are asked to trust and a ranking you are able to verify.

Start with the live index itself, one ranked, sortable and searchable view of all 368 entries, each row carrying its Frontier Score and the three pillars behind it. New this week ranks the papers that entered in the last seven days, and the weekly Frontier Dispatch reviews them. The methodology is public at version 1.3.1, so the way Evidence, Impact and Novelty are computed is not a black box, and the data sources are named individually.

The dataset lives at /data under a CC BY 4.0 license, downloadable as CSV and JSON, free to reuse with attribution. If a figure in this report surprised you, that is where you can pull the record and decide for yourself whether the number holds. To see where any paper stands, inside the index or not, the free paper check shows its field-weighted citation impact, its citations in five open indexes and any retraction on record.

The index is living, not frozen. Its 368 entries cover research published from October 2024 onward, new papers enter daily, and each score is re-checked weekly. Rankings will shift as early signals firm up or fade, and the recency factor will lower today's leaders as they age. Treat any reading, this one included, as a timestamped view of a moving picture, and check the live pages when precision matters.

Those pages tell you how Frontier reads a result. The more portable skill is reading any breakthrough claim, from a press release, a headline or a friend, with the same discipline. Five questions do most of the work.

• Is it corroborated? Look for at least one independent source confirming the core result. A single unreplicated paper is a signal, not a settled fact.
• Does magnitude match stage? An LDL drop of up to 62%, at the highest dose, in a 35-person phase 1 trial is both promising and provisional. Match the strength of a claim to the stage of its evidence.
• Is it new, or confirmatory? Both are valuable, but they are different things. Ask whether the work combines ideas in a new way or confirms an existing one more firmly.
• How far does it reach? A result can be rock-solid yet narrow. Ask who benefits, how soon, and whether the reach is mechanistic or real-world.
• Can you trace every number? If a striking figure has no primary record behind it, treat it as a headline, not a finding, until one appears.

Used together, those questions are the three pillars in plain language: is it real, is it new, does it matter, and can you check it. The index will not tell you a result is mind-blowing; it will tell you whether a result is corroborated, consequential and traceable, or not yet, and it will show its working either way. The aim is not to end argument about the year's science but to ground it in something shared.

Frontier is built by Yuma Heymans (@yumahey), whose background is in large-scale search across roughly a billion online profiles and in consulting rigor, a lineage that shows plainly in the index's insistence that every figure trace to a source. The instinct throughout is the same one that guides good evidence work anywhere: prefer the primary record to the confident summary, and build a system where anyone can check the claim rather than a system that asks to be believed.

• HeroHunt.ai about page: HeroHunt.ai is run by the same founder as Frontier.

So read the dataset, challenge the method, and hold the next "breakthrough of the year" you encounter to the same standard used here: not whether it is admired, but whether it is solid, new, consequential and checkable.

Figures in this report reflect the Frontier Index as of 11 November 2026 (methodology version 1.3.1). The index re-scores entries weekly, so current scores and rankings may differ.

Updated 11 November 2026: this report was first published on 31 August 2026, when the index held 101 entries and its top score was 70, and rewritten on 27 October, when it held 297 entries and its top score was 75. It has now been rewritten on the index as it stood on 11 November: the ranking covers the 295 entries published in 2026, and every count, average, table and chart is recomputed from the live index.

Correction, 7 October 2026: the August edition originally described parts of the method the score does not compute: replication traces, admission of only corroborated work, a reference-graph measure of novelty, and clinical reach under Impact. Those descriptions were corrected that day to match the method.