RLMF
 "Molecular Self-Play" (July 2026), the company's animation of a pore that traps a molecule, reads it repeatedly and releases it, steered by the measured current | |
| Full name | Reinforcement Learning with Molecular Feedback |
|---|---|
| Variant | Reinforcement Learning from Molecular Feedback |
| Proposed by | Molecular Reality Corporation |
| Public since | June 29, 2026 |
| Reward source | The instrument |
| Status | "a hypothesis" |
RLMF, Reinforcement Learning with Molecular Feedback, is a proposal by Molecular Reality Corporation for training artificial intelligence systems with rewards that come from measurements made by molecular sensors. The name adapts reinforcement learning from human feedback (RLHF), in which a model learns from people's ratings of its output. The company's web copy introduces it in one sentence: "Reinforcement Learning with Molecular Feedback (RLMF) is the obvious next step as AIs require more and more real biological data to solve diseases."[1]
The proposal has two parts. In the larger one, AI models of biology would be trained against data from the company's planned network of sensors. In the more specific one, a model would learn to control a solid-state nanopore instrument, choosing voltages and pore geometries, with rewards computed from the instrument's own signals. The company's guide is explicit about its status: "Both are proposals, not products, and I want that said first", and of RLMF, "It is a hypothesis. We are collecting the data that would test it."[2]
Name
Company materials expand the acronym two ways. The public web copy and the guide say "with Molecular Feedback"; other company documents say "from Molecular Feedback", the preposition RLHF uses.[1][3]
A tab labeled RLMF was added to the company's website on June 29, 2026.[3] On June 30, 2026, researchers at Yale University and Google Research posted a preprint introducing "reinforcement learning with metacognitive feedback (RLMF)", in which a language model's rankings are refined according to how well it judges its own performance.[4] The two proposals share nothing but the letters.
Background
RLHF trains a reward model on human judgments, then trains a policy to score well against it. The method was demonstrated on simulated robots and video games in 2017, using people's choices between pairs of short clips,[5] and was applied to instruction-following language models in 2022.[6] Several later methods keep the reinforcement learning and replace the human. A 2025 preprint from a protein-engineering group called its version RLXF, "Reinforcement Learning from eXperimental Feedback", and used laboratory measurements of protein function as the reward.[7] Periodic Labs, founded in 2025 to search for new materials with automated laboratories, describes its approach the same way; its co-founder Liam Fedus said in October 2025 that "our reinforcement learning environments literally derive from the environment, from our physical labs."[8]
Molecular Reality's version applies the substitution to molecules. Its web copy argues that "Even the most intelligent AI systems will need to somehow measure the world of molecules" and that "it will take a whole new level of molecular sensing infrastructure to create the data needed."[1] The same argument underlies the company's tagline, "everybody's racing to build the brain, we're building the eye", in which "the brain" means AI laboratories working on biology and materials, Periodic Labs among them, and "the eye" means sensors.[9]
Origins in the company
An earlier form of the idea appears in the company's December 2022 technical overview, which planned to "adjust the game mechanics to encourage enough data gathering to train advanced deep reinforcement learning models that will learn to optimize translocation control using combinations of ping-pong, sliding occlusion, different pores, different salt and buffer conditions, and several other parameters."[10] In that plan the data would come from volunteers playing games on home instruments, an approach the company calls experigaming.
By 2026 the inspiration named in company study materials was self-play, in which a game-playing program generates its own training data by playing against itself, as AlphaZero did for chess and Go. A white paper of July 2026 titles its section on control "the pore as a game the model learns to play", on the reasoning that an instrument whose controls are moves can produce its own labeled examples.[3] On July 24, 2026 the company published an animation called Molecular Self-Play, captioned "TRAP · SENSE · RELEASE · REPEAT": "A programmable solid-state pore traps a single construct, ping-pongs the analyte through the sensing zone to tomographically sub-sample different conformations, then releases it".[11]
Proposed design
Rewards
A reward based on the identity of the molecule would be circular, since the identity is what the instrument is trying to learn. Company documents therefore propose rewards "measured by the instrument itself, never human annotation": the rate at which molecules are captured, the agreement between repeated reads of one molecule, the signal-to-noise ratio, and the information a measurement adds.[3] The last is the quantity that statistical theory recommends for choosing experiments; it goes back to a 1956 paper by Dennis Lindley on measuring the information an experiment provides.[12] Rewards that do not depend on identity are also meant to resist reward hacking, in which a model scores well without doing what was intended.[3]
Ground truth would be added in stages, which the documents describe as a ladder: weak labels from operators' notes, then samples of known composition, then engineered DNA constructs carrying barcodes, a process summarized as "using labels to get free of labels".[3]
Control
The proposed controller works at three speeds. Fixed logic in the instrument's firmware reacts within microseconds, a small model on the device adjusts thresholds within milliseconds, and a larger model sets the strategy over seconds or minutes. "The slow model never actuates the piezo at microsecond speed; it sets the policy."[3] The actions available are the instrument's settings: the voltage, its timing, the salt and, in the company's sliding occlusion designs, the shape of the pore.
The company's public essay The Nanopore World Model limits the claim. Holding and rereading a single molecule works only when the molecule is confined, by tethering, by two pores pulling against each other, or in a tiny bridge between pores; a free small molecule diffuses away in microseconds. Elsewhere the controller would choose conditions for the next population of molecules, a regime the essay summarizes as "the distribution confesses, not the molecule."[13]
Related work
Nanopore instruments already make decisions in real time. In 2009 a group at the University of California, Santa Cruz used programmable hardware to change the voltage within a fraction of a millisecond of detecting a DNA-enzyme complex.[14] Commercial sequencers can reject unwanted DNA by reversing the voltage once the first stretch of a molecule has been read, a method published in 2016,[15] and a 2023 method chooses which molecules to keep by Bayesian experimental design.[16] In 2025 a study of solid-state pores combined recordings at several fixed voltages with machine learning to tell apart proteins in mixtures; the voltages were set in advance.[17] A literature check by the company in October 2026, covering work from 2015 on, found no published case of a learned policy controlling the voltage or geometry of a nanopore in real time.[3]
Status
No RLMF experiment has been reported. The company's machine-learning results so far come from supervised classification: a Conformer network that reached 85.4 percent accuracy on held-out experiments from a public DNA-barcode dataset, against 81.0 percent for a retrained convolutional network, a result described as unpublished.[3] The guide describes the company's present contribution as data collection.[2]
See also
References
- ^ a b c "RLMF". molecularreality.com. Molecular Reality Corporation. June 29, 2026.
- ^ a b "Keltar's Field Guide to Molecular Reality". Molecular Reality Corporation. August 2026.
- ^ a b c d e f g h i "Company records". Molecular Reality Corporation. October 2026.
- ^ Liu, Gabrielle Kaili-May (June 30, 2026). "Reinforcement Learning with Metacognitive Feedback Elicits Faithful Uncertainty Expression in LLMs". arXiv.
- ^ Christiano, Paul (June 12, 2017). "Deep reinforcement learning from human preferences". arXiv.
- ^ Ouyang, Long (March 4, 2022). "Training language models to follow instructions with human feedback". arXiv.
- ^ Blalock, Nathaniel (May 8, 2025). "Functional alignment of protein language models via reinforcement learning". bioRxiv.
- ^ "Synthesis Superintelligence: from Semiconductors to Superconductors". Latent Space. October 8, 2025.
- ^ Dev Okonkwo-Hale (August 21, 2026). "The Startup That Says Everyone Else Is Building the Brain". The Picoampere.
- ^ Kemmishtree (March 28, 2026). "2022 sci-tech overview". EQBio-OS.
- ^ "Molecular Self-Play". Molecular Reality Corporation. July 24, 2026.
- ^ Lindley, D. V. (1956). "On a measure of the information provided by an experiment". The Annals of Mathematical Statistics 27 (4): 986-1005. doi:10.1214/aoms/1177728069.
- ^ "Keltar's Critique v2: The Nanopore World Model (Hardened)". Molecular Reality Corporation. August 2026.
- ^ Wilson, Noah A. (2009). "Electronic control of DNA polymerase binding and unbinding to single DNA molecules". ACS Nano 3 (4): 995-1003. doi:10.1021/nn9000897.
- ^ Loose, Matthew (2016). "Real-time selective sequencing using nanopore technology". Nature Methods 13: 751-754. doi:10.1038/nmeth.3930.
- ^ Weilguny, Lukas (2023). "Dynamic, adaptive sampling during nanopore sequencing using Bayesian experimental design". Nature Biotechnology 41: 1018-1025. doi:10.1038/s41587-022-01580-z.
- ^ Akita (2025). "Voltage-matrix nanopore profiling for the discrimination of protein mixtures". Chemical Science 16: 18607-18615. doi:10.1039/d5sc05182g.