https://arxiv.org/abs/2602.08079
The question I can’t stop poking: which of these can we write down in engineering terms without slipping into vibes? “Signaling” and “goal-level control” are very real in tissues. What’s the smallest synthetic analogue that could actually scale?
I read it as a gentle jab at the default AI recipe: train a huge model, bolt on tools, cross fingers that generalization appears. Biology leans harder on agents that can build and maintain themselves while staying tightly coupled to the world that’s pushing back.
Design principles they emphasize (the biology route): multiscale autonomy; growth via self-assembly of active parts; constant reconstruction (repair/rewiring); letting embodiment + physics do some of the work; dense signaling so self-organization can coexist with top-down goals.
Their main image is a “cognitive light cone”: how far an agent can sense, predict, and steer across space, time, and scale. Evolution, in this telling, mostly expands that cone—more distant goals, longer horizons, tighter loops of control.
They borrow William James’ framing: intelligence as “fixed goals, variable means.” I like it as a handhold. It points straight at generalization: reaching the same outcome in a situation you didn’t rehearse, instead of replaying a polished policy.
A lot of biology’s “intelligence” shows up before anything like a brain: cells negotiating, wounds closing, embryos turning a noisy soup of parts into a reliable body. This paper argues that if we want robust machine intelligence, we should copy more of that— not just neurons.
Two choices that caught my attention: (1) trained retrievers + rerankers—so it’s not just “vector search and hope.” (2) an iterative self-feedback loop that re-retrieves + rewrites to improve factuality, coverage, and citation accuracy. Same loop makes synthetic data to train a compact 8B model.
Their fix is almost boringly reasonable: first retrieve relevant passages from a big open corpus, then write an answer that stays tethered to those passages. They build OpenScholar on a 45M-paper datastore with 236M passage embeddings.
If you ask a general LLM to “summarize the recent literature” and tack on citations, it’ll often confidently hand you references that… don’t exist. This paper actually measures it: for recent, cross-field queries, GPT-4o’s citations were fabricated ~78–90% of the time.
No benchmark is going to bottle end-to-end discovery. But ones like this can keep us honest about whether “agents” are improving at the workflows researchers actually run, not just test-taking. Link: https://drive.google.com/file/d/1BV5UtmBRdpbQoz9jC1AuUF8WUTRQMqK_/view
They report a real difficulty jump vs the original LAB-Bench: model accuracies drop in a model-specific way, but roughly ~26–46% across subtasks. Even as models get better, more lifelike framing keeps finding the soft spots.
What caught my attention is how unglamorous this is—in a good way. The hard part is often: find the right document, pull the one detail that matters, and resist treating a shaky source like gospel. That’s closer to “useful” than trivia.
The vibe is: fewer multiple-choice guardrails, more open-response + retrieval. Instead of “here’s a figure, answer X,” it sticks figures/tables back into paper-like context. Also tasks on patents, clinical trials, and judging source quality.
LABBench2 is trying to grade that messier layer. ~1,900 biology-flavored tasks spanning literature retrieval/understanding, data access, protocol troubleshooting, molecular biology help, and experiment planning.
A lot of “AI for science” eval still sounds like a school exam: answer the question, show your work. Real research is messier—dig up the right papers, read figures in context, debug a protocol that’s failing, pull sequences, plan experiments, defend the choices.
Fundraising + product announcement (seed $13.5M, Biomni Lab preview): https://phylo.bio/blog/company-fundraising-announcement
What I keep coming back to is the interaction design: steerable agents that can execute real workflows (tools, pipelines, maybe lab robotics) while keeping provenance, citations, and reproducibility intact. That’s a brutal constraint set. Also the right one.
Two things I’d want nailed down (maybe in follow-ups): what “rigorous, safe, reliable” looks like in evals, and what failure modes look like in the wild. Quietly wrong analyses are the nightmare—worse than slow, because you trust them.
They say Biomni (the open-source agent) has been adopted by tens of thousands of scientists across 7,000+ labs/orgs. That catches my eye. Scientists are famously allergic to shiny tools—unless the tool reliably buys them time back on tasks that actually hurt.
Phylo is pitching Biomni Lab as that workspace: one place to orchestrate agents across data modalities and drug discovery stages, with 300+ databases/tools wired in. The vibe is less “chatbot” and more “the OS you live in during real research days.”
A lot of biology is still “glue work.” Ten tabs open, CSVs shuttled around like buckets in a leak, half-manual analyses, scripts you swear you’ll clean up later, tables reformatted… again. The claim here: an “integrated biology environment” + agents becomes the default interface.
Tactile Robotics: Past and Future (Nathan F. Lepora) https://arxiv.org/pdf/2512.01106
If tactile robotics really matures after 2025, the payoff isn’t just nicer grippers. It becomes a testbed for sensorimotor control (what does a policy actually need to know, moment to moment?) and for telepresence where touch feedback isn’t a gimmick.
A note of skepticism runs through it (and through me): tactile sensing has been “a few years away” for decades. The review frames that as recurring bottlenecks—great demos that don’t survive the messy, repetitive reality of deployment. The history makes the loop hard to unsee.
The post-2010 “expansion and diversification” gets broken into categories that feel very familiar if you’ve been around labs: e-skins, tactile robot hands, vision-based tactile sensors (camera + squishy elastomer), soft/biomimetic touch, and the “tactile Internet” (touch over networks).
I like the constraint: mostly reviews only (except pre-1980). It’s a way of sampling expert consensus without pretending you read an infinite pile of one-off papers. Also: a field’s memory often lives in surveys, not in the headlines.
This review tries a simple forecasting move: listen to what ~150 review papers have been saying for ~50 years, then sketch “generations” of tactile robotics: origins (1965–79), foundations (1980–94), a tactile winter (1995–2009), and the post-2010 expansion.
Touch feels like the missing sense in a lot of “dexterous” robotics. You can watch contact, sure—but you can’t reliably feel slip, pressure spreading across a fingertip, shear, or tiny surface texture. That’s how human hands modulate force without crushing or dropping.
From sequence to function: bridging single-molecule kinetics and molecular diversity (Science) https://www.science.org/doi/10.1126/science.adv4503
