Working technical report. The analysis code is public; the original evaluation rows and result traces are not bundled.

Language is an instruction

Arrival is one of my favorite movies. The film turns the Sapir-Whorf hypothesis, or linguistic relativity, into a compelling story where Louise Banks learns the nonlinear written language of the Heptapods, and the language changes how she perceives time.

Language shapes what a speaker expects another person to infer without saying, and what counts as a natural, respectful, or complete answer.

While I’m not sure that I’ll be talking to Heptapods anytime soon, LLMs give me a narrower version of that question. The language of a prompt can change the behavior a model produces. Systems like TranslateGemma sharpen translation quality through supervised fine-tuning, reward models, and human evaluation. But translation quality is only the first step in understanding language.

I wanted to understand how a model follows a language requirement while preserving meaning, maintaining entities, following format, and stopping at the right boundary. These constraints can pull against each other.

I use Japanese and Korean as the core languages and test Cohere’s Tiny Aya family. Their scripts, word structures, and segmentation conventions let me probe constraints an English-only benchmark would miss.

Where does a verifier-valid continuation become available inside Tiny Aya, and what changes between Base and the post-trained checkpoints before decoding?

Core Evaluation

I organized the evaluation around five questions.

Component Question Evidence used here
Meaning preservation Did the model retain the semantic content? Are same-meaning Japanese/Korean prompts close internally? Do outputs resemble references?
Target-language identification Did it know the requested output language and script? Can a simple readout tell which target language is requested? Does the model put probability mass on the right script?
Surface realization Did the right next token win the final branch? Logit-lens rank of valid vs greedy tokens through layers; final branch margin
Format/entity fidelity Did it preserve business constraints? Numbers, dates, URLs, entities, quoted spans, JSON/schema validity
Stop/continuation control Did it stop at the valid answer boundary? End-token scores, prefix-valid failures, overgeneration traces

Why Tiny Aya

Tiny Aya contains a base model and post-trained variants with shared pretraining.

Model Role in this study
tiny-aya-base the pretrained multilingual model, before any instruction tuning
tiny-aya-global the globally balanced instruction-tuned variant
tiny-aya-water a region-specialized merge whose cluster includes Japanese and Korean

The Tiny Aya report frames these as compact multilingual models (3.35B parameters, 36 layers) trained on 70 languages and refined through region-aware posttraining.

The behavioral evaluation uses a small, text-only subset of enterprise-shaped tasks drawn from public datasets.

Source Rows Role
FLORES+ 24 Japanese/Korean translation calibration
Marco-MIF 30 target-language and instruction/format control
DataPilot Japanese Function Calling 10 Japanese tool/schema fidelity
Korean law dataset 10 Korean legal document schema transform
Korean legal QA 10 grounded Korean legal adequacy

Core experiments

The initial sampling and verification experiments used the same 84-row test set. Tracing then examined selected continuations from those runs; the later causal experiments used separate subsets.

  • Sampling: generate 16 outputs per row and ask whether valid behavior exists in the model’s distribution.
  • Verification: score each candidate with a layered verifier built from meaning preservation, target-language control, format/entity retention, and stopping.
  • Internal tracing: compare the valid sampled path against the default path to see where the valid continuation appears, loses, or disappears inside the model.

For sampling, I ask whether the model’s default decoding or likelihood ranking actually selects the valid behavior.I use “greedy” to mean the default highest-probability decoding path. In theory, greedy decoding always chooses the top next token. In practice, modern LLM inference can still be nondeterministic because batching, kernel choices, and floating-point reduction order can slightly change logits. See Thinking Machines Lab, Defeating Nondeterminism in LLM Inference.

I use valid to mean that an output passes the specified deterministic checks. These include reference similarity where available, target script, JSON validity, quoted spans, numbers, dates, URLs, entities, legal references, and answer boundaries. Passing them does not establish human semantic, cultural, or legal adequacy.

  • Greedy valid: whether the default output passed the verifier.
  • Any valid at N=16: whether any sampled candidate passed.
  • Max likelihood at N=16: whether the model’s own likelihood ranking selected a valid candidate.
Model Greedy valid Any valid at N=16 Max likelihood at N=16
Base 24/84 46/84 6/84
Global 45/84 60/84 35/84
Water 43/84 60/84 38/84

Each cell counts rows with a verifier-valid answer out of the same 84 tasks. “Any valid” measures candidate availability, using 16 samples; it is an oracle over the sampled set, not a deployable selection result.

The original report also recorded verifier-plus-likelihood selection at 37/84, 60/84, and 59/84 for Base, Global, and Water. Those counts cannot be interpreted as hard filtering by the same validity predicate over the same candidate sets. Such a filter would recover every any-valid row. The released code does not include the original selector or its candidate records, so I exclude that comparison from the evidence for recoverable validity.

Valid behavior exists in the model’s distribution, but default decoding and sequence likelihood do not reliably pick it. In this evaluation, maximum-likelihood selection underperformed greedy decoding for all three checkpoints.

Why does the valid path lose?

Related evaluations of multilingual constraints

Related evaluations test the gaps between fluent language and instruction-following.

  • Cultural nuance benchmarks find that grammatical adequacy is not the same as cultural resonance. Idioms, puns, holidays, and embedded cultural concepts remain hard even when the output looks fluent.
  • Language and culture are entangled: changing the language of a prompt can change the cultural context the model uses, and low-resource languages often receive lower-quality open-ended answers.
  • Multilingual instruction-following is not just English instruction-following translated. M-IFEval shows wide variation across languages and instruction types, while Marco-Bench-MIF finds high/low-resource gaps, script-specific challenges, and failures from machine-translated evaluation data.
  • Translation can still hallucinate. HalloMTBench separates failures where the model detaches from the instruction from failures where it detaches from the source, and multilingual hallucination work shows that hallucination measurement itself is still often English-centric (Islam et al.).

Looking inside the model

To understand how these models work, I measured seven internal or behavioral signals across Tiny Aya Base, Global, and Water.An internal signal is a measurement taken from model activations, logits, or hidden states rather than only from the final output string.

  • Meaning alignment: do same-meaning Japanese and Korean prompts land closer internally than shuffled controls?
  • Target-language ID: can a simple readout from the hidden state tell which output language the prompt requests?A readout is a lightweight classifier trained on activations. If it can recover target language from a layer, that layer contains linearly accessible information about the requested language.
  • Script control: does the model put probability mass on tokens from the requested script?
  • Format/entity retention: do outputs preserve exact constraints such as numbers, dates, URLs, entities, quotes, and JSON shape?
  • Valid-token availability: does the valid next token become rank-available before final decoding?
  • Branch competition: does the valid continuation beat the default continuation at the final branch?
  • Stop readiness: does the model assign stronger end-of-answer probability to valid endings than greedy-invalid endings?EOS, short for end-of-sequence, is the special token a model emits to signal that the answer is finished. A confident EOS at the right position is what stops generation cleanly.

The measurements sample different parts of the task. A language readout establishes that information is accessible to a classifier. A logit lens projects an intermediate state into token scores. Neither alone identifies the computation the model uses to produce a complete answer.

Related work motivates separating these measurements. Wendler et al. argue that Llama-family models route through an English-like intermediate representation, and Schut et al. find that multilingual LLMs can make key decisions in a representation space closest to English. Paths Not Taken decomposes multilingual factual recall into recall and target-language realization. Translation-mechanism work such as Exploring the Translation Mechanism of LLMs also distinguishes source features, intermediate representations, and target-language realization.

Tracing

Once valid sampled outputs existed, I traced the branch point.

A branch point is the first token where the valid sampled output and the default output diverge. In one path, the model moves toward the valid answer. In the other, it moves toward the default answer that fails the verifier.This is a trajectory comparison. I compare the hidden-state and logit behavior along a valid sampled continuation against the behavior along the default continuation.

At each branch point, I asked three questions.

  • Does the valid token ever beat the default token in any layer?
  • Under teacher forcing,Teacher forcing means scoring a known continuation by feeding the model the correct previous tokens at each step. It lets us ask whether the model can support a trajectory even if it would not choose that trajectory on its own. does the model score the valid continuation coherently token by token?
  • By the final layer, which token wins?
Mean final valid-minus-default token log-probability margins are negative for all three selected subsets: Base minus 1.77 nats on 21 cases, Global minus 1.05 on 15, and Water minus 1.70 on 17.
Reported aggregate means from the branch table below. Negative values favor the default token. Each checkpoint uses a different selected subset; no uncertainty estimates are available. This is not a layerwise trajectory.
Model Branch cases Valid leads in some layer Median best layer Wins at final Mean final margin (nats)
Base 21 20 17 5/21 -1.77
Water 17 16 18 1/17 -1.70
Global 15 14 16 2/15 -1.05

These are selected branch comparisons, not all 84 rows. The final margin is log P(valid token) − log P(default token); positive favors the valid token, negative favors the default. Layer positions are aggregate summaries, not a single trajectory.

The valid token often leads under the intermediate-layer readout, then loses at final decoding. In Global and Water, the valid token sits near the top of the distribution at decoding (median rank 2-3 against the default’s median rank 1). But the final branch margin still favors the default invalid continuation.

Sampling has already established that these models can produce a passing answer on these rows. The branch trace narrows the question to why that continuation loses under default decoding.

Some failures are even simpler. In one Base case, the valid answer was a prefix of the default answer. The model had already said enough, then continued into invalid output.

The branch and stopping failures give me specific decisions to intervene on.

What changes with post-training

The Tiny Aya report makes the comparison useful because Base and the post-trained models sit on different sides of the training process. Base is the pretrained multilingual model, trained for 6T tokens across 70 languages, code, and a multilingual cooldown mix. Global is what happens when that substrate is shaped with instruction data, translation expansion, cultural adaptation, synthetic teacher completions, and a light preference-tuning step that anchors identity and safety. Water is built on top of Global by training a regional checkpoint (a cluster that includes Japanese and Korean) and merging it back into Global via SimMerge, a checkpoint-merging procedure.

The behavioral comparison favors Global and Water on this packet.

  • Greedy valid rows rise from Base 24/84 to Global 45/84 and Water 43/84.
  • Exact invariant retention rises from Base 0.753 to Global 0.904 and Water 0.911.
  • Valid endings have higher EOS logprob than greedy-invalid endings across all three models, Base +1.55, Global +1.82, Water +1.03.

The invariant-retention scores summarize exact constraint preservation. The EOS differences compare valid and greedy-invalid endings; their sample counts were not stated in the published summary. These descriptive results accompany the clearer 84-row improvement in greedy validity.

I then tested whether moving a late residual state between checkpoints would change the next-token decision or repair the full answer.

What the interventions changed

The follow-up used a separate triangulation corpus with 14 paired source items per language pair. I tested four translation directions, Japanese-English, Korean-English, Arabic-English, and Japanese-Chinese. Across all three Tiny Aya models, all 12 model-pair conditions showed a same-meaning signal peaking around layers 27-30.The same-meaning signal is the cosine gap between paired prompts with the same meaning and rotated prompts with different meanings. It asks whether meaning is preserved beyond surface language overlap.

The instruction-following analysis used eight content-parallel Japanese/Korean pairs as its primary subset, with 17 pairs in the broader analysis. The single-language controls used 12 rows per task family. These follow-up populations differ from the initial 84-row packet.

Task Representational observation
Translation The same-meaning signal was present in Base and stronger in Global and Water.
Paired instruction-following Base had no clean late peak under this measurement; Global and Water peaked at layer 26.
Single-language controls Checkpoint differences on function calling, Korean law, and Korean legal QA concentrated around layers 13–21.

A late signal became measurable after post-training on paired instruction-following. That observation does not by itself show that training installed a new circuit.

Residual patching replaces a recipient model’s hidden state at a chosen position and layer with a donor state from another checkpoint. I tested both directions where available. A change establishes sensitivity to that intervention, while full-answer repair is a separate outcome.

Task and patch Layer Reported effect Full-answer result
Translation, Global → Base 34 +2.79 nats in the Global-token minus Base-token branch margin Not established by this score
Translation, Base → Global 34 −5.55 nats in the same branch margin Not established by this score
Korean law, direction not reported 34 −3.38 nats in continuation log probability No full-answer repair count reported
Marco-MIF, Global → Base 26 +2.25 nats in first-token log probability 0/8 reported full-outcome cases repaired

Translation uses the first token where the checkpoints diverge; positive shifts favor Global’s token. This differs from the earlier valid-minus-default margin. Korean law scores a teacher-forced continuation; Marco-MIF scores its target continuation’s first token. Per-effect sample counts and uncertainty estimates were not reported.

The released patching code defines these metrics and processes both language sides of eight Marco-MIF pairs (16 prompts). Its row count should not be substituted for the original eight reported full-outcome cases.

The Marco-MIF patch shifted the first-token score, but the effect did not persist through the answer in the reported test. I could move that local decision without repairing any of the eight full outcomes.

The singular value decomposition (SVD) of the Global-minus-Base weight change at layer 34 was broad. It took 1564 of 2048 directions to capture 90% of the change. A small SVD edit moved Base toward Global by +0.22 chrF, a character-based similarity score measured against Global’s continuation rather than a human reference. Weight-difference rank alone does not establish the dimensionality of the behaviorally relevant change.

A failed feature diagnostic

The sparse autoencoder (SAE) analysis initially seemed to support a distributed explanation. It learned usable reconstructions but returned a top discrimination score of only 9.5e-07, and the Base and Global dictionaries aligned at cosine similarity 0.999.

On rereading the released analysis, I found a problem with that interpretation. The statistic averages Japanese and Korean feature activations within each pair, then compares the overall mean for matched pairs against the mean after permuting the Korean rows. Permutation preserves that mean. The score is zero by construction, apart from floating-point error, regardless of whether the features encode useful information.

The SAE training code also deliberately shares initialization between runs. The high dictionary alignment needs to be evaluated against that starting point. Neither statistic establishes the absence of a sparse multilingual feature, and I withdraw that inference.

Parameter-level approaches such as Goodfire’s adVersarial Parameter Decomposition (VPD) remain relevant to the question. This experiment has not yet isolated a comparable explanation of the weight changes.

The strongest result is narrower. On this packet, post-trained checkpoints produce more verifier-valid answers under greedy decoding. Selected branch traces show passing continuations losing at final decoding. In the eight reported Marco-MIF outcomes, a single residual patch changed the next-token score without repairing the answer. The open question is what intervention can carry that change through the full continuation.