Experiment 25 — activation patching on learned chunk representations
Are the units a dynamic chunker invents genuinely functional computational units, or compression artefacts?
Run: python experiments/exp25_chunk_patching.py (~8 min, 1 GPU).
Artifacts: patches.csv (52,025 patches), summary.json; site-redraw replicate in patches_siteseed1.csv / summary_siteseed1.json.
Adapter: interp/runner.py. Estimators: interp/patching.py, unmodified.
Headline
No. At this scale the learned chunks are not privileged loci of information. Replacing a whole chunk state at its boundaries and replacing the same number of bytes starting mid-chunk are, within across-seed spread, the same intervention. Aligned-minus-misaligned averaged over depth:
| condition | aligned (damage, nats/byte) | misaligned | contrast (A−M) | contrast − Gaussian null |
|---|---|---|---|---|
| A (no parity) | −0.0091 | −0.0114 | +0.0023 ± 0.0039 | +0.0028 ± 0.0049 |
| B (parity, per-sent) | −0.0071 | −0.0184 | +0.0114 ± 0.0072 | +0.0070 ± 0.0070 |
| Apc (no parity, per-cp) | −0.0094 | −0.0081 | −0.0013 ± 0.0020 | +0.0024 ± 0.0036 |
| Bpc (parity, per-cp) | −0.0085 | −0.0103 | +0.0017 ± 0.0064 | +0.0009 ± 0.0019 |
± is the spread across the 3 seeds, never pooled over items. Every value is a mean over
layers 0–6; the per-layer table is in summary.json under results.mono.
0/28 (family × layer) cells have |mean| > 2 across-seed SD under an independent draw of patch
sites; 1/28 do under the first draw (B/L0), and that one cell does not survive the redraw. Of 84
item-level permutation tests (one per family × layer × seed), 6.0% fall below p = 0.05 — the rate
you get by construction.
Why the effects are this small: the chunk pathway itself
Without a positive control the number "0.01 nats" is uninterpretable. Two ablations, whole-sentence, nats/byte, averaged over the 12 checkpoints:
| intervention | Δ NLL |
|---|---|
| base NLL | 1.417 |
| zero the entire main-network contribution to the decoder | +0.344 (24% of the loss) |
| replace every chunk state (26 of them) with an unrelated sentence's | +0.243 |
| that ceiling divided by the mean chunk count | +0.0093 |
| replace ONE chunk state (aligned, layer 0) | +0.0176 |
The chunk pathway is genuinely load-bearing — deleting it costs a quarter of the model's predictive performance. But a single chunk's causal contribution is within ~2× of the aggregate effect divided by the number of chunks. The information is additive and diffuse across chunks, not concentrated in the unit that the router carved out. That is the negative result, and it is the same result whether the edit respects a boundary or not.
Depth
Effect magnitude decays monotonically with depth in every family — layer 0 (states entering the main net) ≈ 0.018 nats damage, layer 6 (final main block) ≈ 0.004 — i.e. a chunk's identity is most consequential where it enters the stack and is progressively mixed away. The aligned-vs-misaligned contrast shows no depth structure: it is flat and inside seed spread at every layer.
Limitation. Only ckpt_last.pt (step 6000) was retained for these 27 runs, so this is depth at
the FINAL step only. The layer × training-time version — where the interesting question (when do
units become causally functional?) lives — needs intermediate weights that do not exist and would
require re-running training with keep_ckpt on.
A vs B
Condition B (parity-trained, per-sentence β) is the only family with a consistently positive contrast, present at all 7 layers and in both site draws (+0.0114 ± 0.0072 and +0.0098 ± 0.0114). Its misaligned edits do ~2.6× the damage of its aligned edits. This is suggestive and it is not significant: the across-seed SD is comparable to the mean, it does not survive subtraction of the statistics-matched Gaussian null in the replicate (+0.0025 ± 0.0016), and the matched per-codepoint pair (Apc vs Bpc) does not reproduce the ordering (−0.0013 vs +0.0017). With 3 seeds it is an anecdote. Reported because the pre-registration says both answers are results, not because it supports the parity hypothesis.
Cross-lingual (secondary)
patching.ChunkPatchStudy run unchanged, donor = English, recipients de / hi / zh / ta, layer 0,
content-matched (parallel FLORES+ row) against the content-mismatched-donor null:
- aligned-minus-null across all 16 (family × language) cells: mean −0.0006, range −0.0083…+0.0072, mean across-seed SD 0.0059. 1/16 cells exceed 2 SD — chance.
- Donating an English chunk into a Hindi or Tamil forward pass does nothing distinguishable from donating an unrelated English chunk.
Caveat inherent to the design: the donor chunk is taken at the same chunk index, not the same
content, so cross-lingual chunk k is not the translation of recipient chunk k. This is
patching.py's pre-registered design and it makes the cross-lingual leg a weak test.
Controls that passed
- Identity (the load-bearing assertion). Patching a chunk with its own vector is a bit-exact
no-op —
max |Δ NLL| = 0.0— via both injection paths, at layers 0, 3 and 6, on all 12 checkpoints. The bytes path only degenerates to the chunks path when the byte range coincides with a chunk span, so this tests the overlap arithmetic itself. The driver raises if it is non-zero. - Forward fidelity.
runner._forward(which threads an injection point through the main stack) agrees withHNet.forwardtomax |Δ logits| = 0.0when unpatched, so the depth sweep cannot silently diverge from the trained model's own computation. - Statistics-matched Gaussian donor (right per-dimension mean/SD at that layer, no content),
run in both aligned and misaligned form, so the reported contrast has its own null contrast.
Overall
contrast − null= +0.0033, mean |value| 0.0034, mean across-seed SD 0.0055. - Perturbation mass. A misaligned byte range does not carry the same total overlap weight as
the chunk it displaces: mean mass 1.18 vs 1.00. Restricting to misaligned patches with mass in
[0.9, 1.1] does not change any conclusion (mass-matched contrasts in
summary.json). - Site-redraw replicate. Independent draw of which chunk is patched: contrasts correlate r = 0.64 across the 28 cells with 89% sign agreement and mean |difference| 0.0031 — comparable to the effects themselves. Item-sampling noise is not negligible at this effect size.
What was built
interp/runner.py — the adapter interp/patching.py asks for, closing the [GAP] flagged in
interp/README.md §5.3:
encode(lang, sent_idx)/nll(lang, sent_idx, patch=None)againstscratch/ckpt/<run>/ckpt_last.pton the trainer's byte-identical FLORES+ probe (probe/probe_text.jsonl), one sentence per forward pass, fp32, no padding.- Both injection paths:
{"chunks": {j: v}}(aligned) and{"bytes": ((b0,b1), v)}(misaligned), the latter viamodel.apply_byte_patch, which writes into every overlapping chunk state in proportion to overlap. - Injection at arbitrary depth of the main network (layer 0 = chunk states entering it;
layer l = output of main block l−1), with donor vectors drawn from the donor's forward pass at
the same layer. This is new;
HNet.forwardcould only inject at the main net's input. - Positive-control ablations
{"kill_main": True}and{"all": matrix}.
experiments/exp23_interp_validation.py still passes 49/49 (15.9 s).
Other limitations
- 16.2 M-parameter models, 6000 steps, one hierarchy stage. A causal null at this scale is not a causal null at scale; it is evidence that unit-localisation is not automatic.
- Probe = 96 English FLORES+ sentences with ≥ 6 chunks, truncated at 512 bytes, mean 26 chunks each.
- The architecture caps the achievable effect: the byte-level encoder residual reaches the decoder
directly (
decoder(h + res_gate * up)), so 76% of the model's prediction quality is untouchable by any chunk-state patch. An H-Net without that residual would be a sharper test. - Seed spread is estimated from 3 points, so "± SD" is indicative, not an interval.