By Luxia (@kataletheia), with Claude (Anthropic) · compute and hosting by Anima Labs
This atlas scores 116 models and checkpoints, plus one human reader, on 28 tests of tracking who I and you refer to. Click any cell in the full atlas to read the exact prompt that the model was given.
Base language models acquire the referent of you in five ordered stages. Every base of 0.6B parameters or more trained on trillions of tokens reaches the first three. No public model up to 8B reaches the fifth stage zero-shot (with no worked examples): tracking you after objects change hands. That holds at every token count measured, up to 36T, and at their final checkpoints every one of them has a negative swap shift. From 14B to 1T parameters, stage 5 is weak and nearly flat (Figure 2). In these models, parameter count and depth grow together. kotodama-tsumugi separates the two. It is a 1.4B model run as 72 effective layers, with looped depth and boundary-fitted attention residuals. It reaches stage 3 by 33B tokens (annealed; 40B on the constant-learning-rate trunk), about what OLMo-2 13B needs, where OLMo-2 1B needs about 420B. At 110B tokens its zero-shot dynamic you is not inverted (−.021), where most 1–2B bases invert. In a check of 20 bases, it follows a narrated hand-off better (.336) than every Pythia model up to 6.9B (.275 to .294). For every model of its size, that hand-off is the open step toward stage 5.
The stages appear in the same order in every trajectory measured: OLMo-2 1B, 7B and 13B; SmolLM3 3B; Pythia 410M to 6.9B; kotodama-tsumugi (Part III). Each stage is read with a speaker swap, and the hatched stage is where most bases stop. Who Is You? explains the stages and the swap.
Dynamic you at each model's final checkpoint (OLMo-2 1B: its 4T pre-anneal checkpoint). A positive swap shift means you follows the addressee; a negative one means you resolves to the speaker. Zero-shot (filled) is negative for every model at or below 8B, for token counts up to 36T, and turns positive between 8B and 14B on both Qwen recipes. The same holds with each chat rewritten as a scene (8 of 8 models tested at or below 8B; the highest is +.010, Qwen2.5-7B), so the result does not come from the chat format. In all of these standard-shape transformers, depth grows with size. The chart therefore leaves two routes open for small models: more tokens, and more depth at a fixed size. No model at or below 8B has been trained far enough to rule out the first. From 14B to 1T parameters, stage 5 is weak and nearly flat. Ten of the 11 bases with a public size read +.01 to +.07 zero-shot: +.022 for Qwen3 14B, +.054 for Llama-3.1 405B, +.045 for Kimi-K2 at about 1T. Static you reads +.27 to +.82 in the seven of them where it was measured. In-context (3-shot, hollow) turns positive at smaller sizes and reaches +.24 for Qwen3 14B. The kotodama-tsumugi release (diamond; 1.4B at 110B tokens) reads −.021 zero-shot, not inverted under the gauge rule. Its interval reaches −.013, and the rule calls a reading inverted only when the whole 95% interval lies below −.02. Of the 1–2B public bases whose static you binds, only Llama-3.2-1B (−.022) shares that; the rest invert (−.047 to −.089).
Training tokens until each stage appears, for one data recipe (OLMo-2) at three sizes. Stage 1 costs about the same at 1B, 7B and 13B. Stage 3 takes about ten times fewer tokens at 13B than at 1B. Stage 5 is absent at the last checkpoint of all three (4T to 5T tokens). kotodama-tsumugi, at 1.4B parameters, reaches stage 1 between 3B and 3.75B tokens and stage 3 between 30B and 33B: the 13B model's token cost at a tenth of its size.
The dynamic-you items, rewritten along one axis at a time (accuracy at 0-shot, self and addressee queries pooled; chance ≈ .29; about ±.024 per cell). Narrated removes deixis from the whole scene. Spoken is the gauge item itself. Spoken, transfer narrated keeps the chat but states the hand-off as a stage direction. Spoken + vocative names the addressee in the query. Three layers show. (1) Possession transfer itself. Every 1.4–1.5B model is at or near chance even when the scene is narrated: kotodama-tsumugi through 110B, Pythia-1.4B, OLMo-2 1B at 4T. SmolLM2 1.7B at 11T and Qwen3 1.7B at 36T lift off chance; from 4B it is easy (.57–.80). (2) The transfer as a speech act (“here, take my cup”) is the mid-size wall: models from 4B to 70B that pass the narrated version stay at .29–.41 when it is spoken. Narrating only the hand-off recovers part of that gap, from almost none (Qwen3 4B) to about three quarters (Qwen3 30B-A3B). (3) Naming the addressee does not help: the vocative variant is at or below the spoken one for every public model, and within about .01 of it for kotodama-tsumugi.
Every reachable model against every referent-binding probe, chance-normalized (0 = guessing, 1 = perfect). Click a cell to read the exact text the model was given.
Each lever measured on 1.4B-parameter arms trained for 4B tokens on one data order; one seed per arm except the fitted boundaries (three seeds). Set binding and attribution are capacity probes from the battery. kotodama-tsumugi (Part III) combines the three positive levers in one model.
言霊 kotodama, word-spirit: language as a thing with a life of its own. 紬 tsumugi, spun thread: words woven into more than the sum of their parts. Every neuron of this model, the circuits behind these probes on items with no worked examples, and which parts of them hold when the format changes: Threads.
A 1.42B-parameter model with a looped core, run as 72 effective layers. Its trunk trains at a constant learning rate to 100B tokens, and annealed reads are cooldown branches forked from its checkpoints; the release is the 110B branch from the 100B fork. The architecture and training run are described in Who Is You?, and Table 4 gives the spread between two anneals.
Swap shift against training tokens for the three stage gauges, seed 17's trunk (filled markers, constant LR to 100B) and its annealed branches (hollow), with seed 7 (small grey markers, to 11B). Static I starts to bind at 3.75B tokens and levels off at +.19 to +.31 after about 30B. Static you first dips below zero, the stage where the model treats you as me. It first binds at 33B tokens in an annealed read (+.073) and at 40B on the trunk itself (+.077), and reaches +.235 at the 110B release. The second anneal from the same fork reads +.142 (Table 4 gives the anneal-to-anneal spread). Dynamic you (0-shot) inverts from about 4B, deepest at 11B (−.047), and shallows to −.020 at 88B and −.016 / −.021 at 110B, no longer inverted under the gauge rule. The markers along the top show where OLMo-2 reaches stage 3 at 13B, 7B and 1B parameters.
Every same-harness checkpoint of kotodama-tsumugi and its peers, matched on the x-axis. The white line is tsumugi's constant-LR trunk; hollow markers are its annealed reads, the 110B release among them. Grey lines are public checkpoints of Pythia-1.4B, OLMo-2 1B, OLMo-1 1B and SmolLM2-1.7B. Held-out text is bits per byte, drawn with lower (better) at the top. On the FLOPs axis tsumugi pays for its looped depth (1.13e10 FLOPs per token, about 1.3× the 1–2B bases).
By the release, attribution reads .986, level with the 7–8B bases under the same balanced examples (.981 to 1.00). The table gives accuracy on the binding battery at each annealed branch, scored with one harness (the kotodama binding suite). Each item gets four worked examples, balanced so that none of them shows the item's own answer pattern; deixis-lite has no such condition and is read zero-shot. Chance is in brackets where it is fixed. The bookkeeping probes rise with tokens: attribution from .71 to .99, set binding and dialogue tracking by .13 to .17. Aliased Python lists stays at chance through 33B and lifts only after 55B. The perspective probes stay within .15 of their 5B level. The two seeds differ at 5B (seed 7 forms set binding earlier) and agree within .05 on seven of the ten rows at 11B.
The table has three kinds of column. The stage gauges are swap shifts, and a shift above +.02 counts as binding. The capacity columns follow arXiv 2609.30634. K50 is how many name–task pairs fit in context before recall falls halfway to chance. The last capacity row is the recall of one binding across 256 tokens of filler. The deixis-v2 columns are accuracies (chance ≈ .33). The two 110B columns are two anneals from the same fork with the same training loss. The spread between them (static you .142 vs .235, K50 3.8 vs 4.8, and .025 bits/byte on held-out text) is the measured anneal-to-anneal variance. Every single-anneal number on this page carries that spread, so a difference between neighbouring checkpoints on one axis is not a trend. A second initialization seed (seed 7, same data order) was trained to 11B. deixis v2 has no SmolLM2 or Qwen3 grid. Capacity tracks recipe and tokens more than parameters on this harness (Pythia is flat from 160M to 6.9B), so no parameter-equivalent size is read from it.
Riders v2: boxes go through swaps and moves in narrative or screenplay prose and one box is queried (accuracy; chance .25; about ±.024 per cell). Zero-shot, the annealed kotodama-tsumugi checkpoints from 33B on (.455–.490) sit above every public checkpoint measured on this harness (best: SmolLM2 1.7B at 8.4T, .420). That comparison pools narrative and screenplay prose, and the public pool is three 1–2B models. fixed3 gives every item the same three exemplars. Under fixed3, accuracy falls well below zero-shot for OLMo-2 1B at 4T and for every kotodama-tsumugi checkpoint (−.05 to −.22). The few-shot penalty in the probes run during training came from that setup, not from a failure of in-context learning. The penalty is not specific to these models. On a lighter riders slice, fixed exemplars lower prose accuracy for most of an eight-model panel (chat 8 of 8, screenplay 6 of 8, narrative 5 of 8) and raise code accuracy for all eight. With exemplars drawn per item (fair1 / fair3 / fair6), few-shot helps every public model here (+.06 to +.11 at fair6) and kotodama-tsumugi through 55B (+.02 to +.08). The two 110B anneals score highest zero-shot (fair6: second anneal −.03, release −.045). Object tracking is a separate skill from stage 5: OLMo-2 1B climbs from .230 at 1B tokens to .385 at 4T while its zero-shot dynamic you never binds.