← Research

TALK

Give a language model real cells as RNA tokens and ask what they are. Every answer here is a recorded run of our trained models; nothing runs live.

RNA-LLM · Public PBMC data (Kang et al. 2018) · 256 cells · 3 model versions

Architecture · Figure 1

From RNA to language

Try TALK ↓

RNA → encoder → Q-Former → RNA prefix + question → Qwen3 → answer

Loading the diagram…

Loading the recorded answers…

Cells

Condition

How many

Model

Fine-tuned = supervised training. GRPO and DAPO add 500 steps of reinforcement learning to it.

Prompt and answer

Prompt ·

TALK

Select a gene for its claim and measurements. Colors grade each mention, not detection alone.

What TALK gets

Each dot is one cell’s RNA vector, the input TALK reads. Selected cells are bright.

Measured in these cells

Mean log1p(CP10k) of marker genes. TALK never sees these numbers as text.

Answer evaluation

8 cells per group · 3 models · 8 groups

Cell type

8/8

All three models identified the cell type in all eight groups.

IFN-β stimulation

4/8

Every answer was Control, including the IFN-β-stimulated groups.

Gene descriptions

57.7%

1,122 of 1,946 assessed gene mentions matched the measurements.

157 ungraded mentions excluded

Results and scoring

Cell type / stimulation · correct answers out of 8 groups
Model	1 cell	8 cells	16 cells	32 cells
Fine-tuned	4/8 · 5/8	8/8 · 4/8	8/8 · 4/8	6/8 · 5/8
GRPO	4/8 · 5/8	8/8 · 4/8	8/8 · 4/8	7/8 · 5/8
DAPO	5/8 · 5/8	8/8 · 4/8	7/8 · 4/8	8/8 · 4/8

Gene mentions · 8 cells per group
Model	Matched / assessed	Rate	Ungraded
Fine-tuned	378 / 673	56.2%	35
GRPO	384 / 640	60.0%	35
DAPO	360 / 633	56.9%	87
Total	1,122 / 1,946	57.7%	157

We checked detection and rank against the measurements, and corpus comparisons against the training baseline. Overlapping tied ranks count as a match. For “higher,” we check whether expression exceeds the corpus baseline. For “high in both,” the input must rank in the top 3% or top 20, and the corpus in the top 1%.

The rate is green mentions divided by all assessed mentions (green, red and yellow). Repeated mentions of the same gene are counted separately. The 157 gray mentions are excluded.

The eight groups combine four cell types with Control or IFN-β stimulation. We tested groups of 1, 8, 16 and 32 cells, with one answer per model for each group and input size. Reasoning was enabled and temperature was 0.6. Unfinished answers count as incorrect.

We used public data from Kang et al. (2018). The recorded study metadata lists this study as excluded from training.