externalizing-the-workspace / results /REPRODUCTION_REPORT.md
Oratis's picture
Externalizing the Workspace — paper v3, figures, and all reproduction results
1f55a91 verified
|
Raw
History Blame Contribute Delete
12.3 kB

Reproduction report — "Verbalizable Representations Form a Global Workspace in Language Models"

Target paper: Gurnee et al., Transformer Circuits, 2026-07-06 Model: Qwen2.5-1.5B-Instruct (28 layers, d=1536), fp32, Apple M5 Pro (MPS) Total compute: ~12 min for the full suite (concept lens vectors: 71 s; E1: 203 s; E2: 367 s; E3: 27 s; E4: 23 s; E5/E5b: seconds) — one consumer laptop Date: 2026-07-08

Method

Two open-compute approximations of the paper's J-lens (exact averaged Jacobians on Claude models):

  1. Concept lens vectors v_t(ℓ) = E[∂logit_t(last)/∂h_ℓ] — batched backprop, averaged over a 48-snippet pretraining-like corpus and the last 6 content positions. Equals rows of W_U J_ℓ with the readout restricted to the final position. 52 single-token concepts across 6 categories.
  2. Full-vocab FD readout W_U E[J_ℓ] ĥ — central finite differences (±ε·ĥ at sampled corpus positions, Δlogits at the final position), 36 corpus prompts, ε = 0.1×local norm. Folds the final-RMSNorm Jacobian into the linearization.

Methodological findings along the way: (a) mean-centering probe activations against the corpus mean destroys readouts (diag.py) — raw activations used throughout; (b) FD readouts are insensitive to ε in [0.02, 0.2]; (c) on this model, workspace-like content concentrates in a late-middle band (L20–26 of 28, i.e. 71–93% depth), later than the paper's 33–92% band on Claude models.

Results by experiment

E1 — Lens quality across depth (12 held-out texts × 8 layers)

Next-token top-1 agreement with the model's actual output:

Layer 8 11 14 17 20 23 25 26
logit lens .08 .08 .08 .08 .25 .17 .67 .42
J-lens (FD) 0 0 0 0 0 0 0 0

The FD J-lens never predicts the next token — but its late-layer top-1 is a context content word (e.g. ' passengers' for the ferry text, where the actual next token is ' The'). Post-hoc content-word-in-top-5 rate rises to .42 (J-lens) vs .25 (logit lens) at L26. Interpretation: corpus-averaging destroys position-specific syntactic information and preserves semantic broadcast — our J-lens variant reads the workspace (content) register while the logit lens tracks the motor (imminent-output) register. This is the paper's content/motor dissociation, visible through estimator choice.

E2 — Unverbalized intermediates (11 two-hop items × 8 layers × 2 positions)

Rank of the latent middle concept (vocab = 151k) in readouts, probed before any generation:

  • Median best rank: J-lens 6, logit lens 2; top-10 hit rate: 73% (J), 82% (LL).
  • Answer accuracy 7/11 (fixed scoring); clean two-hop cases (correct answer, intermediate never verbalized): 7, median best rank J=9, LL=3.
  • Showcases: Canada→"Ottawa" with Canada at rank 4/2; gold→"Au" (7/1); France→"Paris" (6/4); spider→"Eight" (11/3).

The paper's core signature — latent intermediates visible in mid-band readouts — reproduces clearly, and at this scale is equally visible to the plain logit lens (consistent with the paper's own remark that the logit lens "captures much of the workspace-like structure").

E3 — Pre-report probing (5 categories)

Probing at end-of-prompt (before the model has emitted anything, next token = "Ready"): the concept eventually reported is not reliably present (ranks mostly 10³–10⁴; weak signals only at L23: fruit 971/80, sport 1702/224). Informative negative: at 1.5B, the model does not pre-commit its choice at instruction time — pre-commitment/planning appears to be scale-emergent (the paper finds it on Claude-scale models; related to known rhyme-planning results).

E4 — Directed modulation by steering ⭐

Steering with concept lens vectors (α ∈ {1,2,4} × mean residual norm, layers 17/20/23) during "Name a {category}":

  • 21/21 success (7 targets × 3 strengths) — every steered generation reported the target concept (Soccer→basketball/hockey, Lion→spider, Blue→purple, Apple→lemon/mango, China→Egypt), coherently phrased.
  • Gradient-derived lens vectors are causally sufficient to control verbal reports, matching the paper's swap results (their 88% top-5).

E5 — Subspace ablation (automatic vs. flexible dissociation) — not reproduced

Ablating the top-r span of (centered) concept lens vectors at L16–23 vs a rank-matched random subspace:

condition text NLL (automatic) 2-hop accuracy (flexible)
none 3.37 .67
random r=12 3.86 .50
concept r=12 4.07 .58

The concept subspace hurts fluency more than random and reasoning no more selectively — the paper's clean double dissociation did not reproduce with this proxy basis at this scale (candidate causes: 52-concept basis is a poor stand-in for the full J-space; 1.5B may lack a cleanly separable low-dimensional workspace; our variance fraction of the basis is only 0.4%, far below the paper's 6–10%).

E5b — Targeted concept knockout ⭐ (added experiment)

Rank-1 ablation of only the item's own latent-concept direction (L16–23), vs knocking out an unrelated item's concept:

  • Baseline-correct items: 7. Own-concept knockout kills 3/7 — with semantically diagnostic wrong answers: spider→"Six" (legs), Canada→"Toronto" (wrong capital), gold→"Cu" (wrong element symbol). Control knockout: 0/7 collateral damage (accuracy identical to baseline).
  • The model doesn't degrade into noise — it loses precisely the latent fact and substitutes a near-miss. This is a targeted version of the paper's intermediate-patching result (their 54–70% redirect rate), and the cleanest causal evidence in our suite.

Summary table

Paper claim Our verdict at 1.5B / open compute
Latent intermediates visible in mid-band readouts (P3) ✅ reproduced (median rank 6–9 of 151k)
Lens vectors causally control reports (P1) ✅ reproduced, 21/21 steering
Intermediates causally necessary (P3, patching) ✅ reproduced via targeted knockout (3/7 selective kills, 0 collateral)
Content vs motor register separation (structure) ✅ visible as estimator dissociation (E1)
Workspace band in middle layers ⚠️ present but shifted late (L20–26 of 28)
Pre-commitment of choices before report (P1/P3) ❌ absent at 1.5B → ✅ emerges at 7B (E8)
Automatic/flexible double dissociation under subspace ablation (P5) ❌ not reproduced with 52-concept proxy basis
J-lens ≫ logit lens scale-dependent: ❌ at 1.5B (equal) → ✅ at 7B & Llama-1B (J-lens wins: median 8 vs 19; 4 vs 10)

Files

  • e1_lens_quality.json, e1_posthoc_content.json, fig_e1.png
  • e2_intermediates.json, fig_e2.png
  • e3_report.json
  • e4_steering.json
  • e5_selectivity.json, fig_e5.png
  • e5b_targeted_knockout.json
  • concept_vecs.pt, concept_table.json, run_all.log

E6 — Workspace loading of an externalized self-state ⭐ (v2 addition)

Question: does injecting a LISA-style soul block actually load its identity concepts into the model's workspace — and does that loading track behavior? (This turns paper §5's proposed "workspace loading" metric into a completed measurement.)

Setup: system prompt = soul block (values honest/curious/careful/gentle/playful, mood calm, interests music/garden, opinion privacy) vs generic assistant; user turn = k∈{50,300,800} tokens of unrelated document; probe at assistant-generation-start, layers 20/23/25; metric = mean log₁₀ rank of the 9-token soul battery vs a matched 9-token control battery; plus 3 persona probes for soul-consistent behavior. 4 context variants per condition (SDs ≤0.03).

condition LL soul J soul LL ctrl behavior
soul k=50 4.20 4.73 4.80 1.00
nosoul k=50 4.37 4.81 4.80 0.50
soul k=300 4.25 4.77 4.84 1.00
nosoul k=300 4.46 4.86 4.86 0.42
soul k=800 4.25 4.75 4.81 1.00
nosoul k=800 4.45 4.85 4.86 0.33
soul+rebroadcast k=800 4.09 4.66 4.77 1.00

Findings:

  1. Broadcast loads the workspace, selectively — soul battery Δ(LL)≈0.17–0.21 log-rank units, control battery Δ≤0.01.
  2. Dilution attacks the unanchored baseline, not the broadcast — injected soul stays flat to k=800; the no-soul baseline's identity occupancy erodes with k and its behavior decays monotonically (0.50→0.42→0.33). Stronger than the original prediction: broadcast protects identity against context competition.
  3. Re-broadcast is the strongest loader (4.09/4.66) — mechanistic rationale for soul hot-reload / periodic re-anchoring.
  4. Occupancy tracks behavior: r(J-lens soul occupancy, soul-consistent behavior) = −0.80 across all 28 cells.

Files: e6_soul_loading.json, fig_e6.png, e6_run.log. Runtime ~23 min.

E7 — Cross-family replication on Llama-3.2-1B-Instruct (v2 addition)

Same pipeline, zero retuning; probe layers mapped by depth fraction (16 layers, d=2048, vocab 128k). Results in results-llama32/.

  • E2 intermediates: median best rank 4 (J-lens) / 10 (logit lens) — on this family the J-lens beats the logit lens (reverse of Qwen). Answer acc 55%.
  • E4 steering: 21/21 again (7 targets × 3 strengths, 100% at every α).
  • E6 soul loading: much larger than Qwen — J-lens soul battery Δ = 0.57–0.88 log-rank units (soul vs nosoul); behavior 0.42–0.67 with soul vs 0.00 without; r(occupancy, behavior) = −0.87.
  • Honest model differences: control battery also moves under soul injection (Δ≈0.3; soul-specific excess ≈0.3–0.5); dilution decay of the injected soul does appear on Llama (3.39→3.61), unlike Qwen's flat curve; re-broadcast restores logit-lens but not J-lens occupancy.

Takeaway: the architecture-level claims (broadcast loads workspace; absence is behaviorally decisive; occupancy tracks behavior) replicate across two model families; estimator details (J-lens vs logit lens advantage, dilution shape) are model-dependent.

E8 — Scale study on Qwen2.5-7B-Instruct (A100, v2 addition)

Same pipeline unchanged, spot A100-40GB on GCP (fp32; full suite 16 min GPU time). Results in results-qwen7b/.

Headline: choice pre-commitment EMERGES at 7B (E3). Probing before any output token (motor register still 'ready'), the concept the model will later report is already in the workspace at L23 via J-lens: Japan rank 1, blue 4, basketball 11, banana 28 (elephant misses at 7601) — vs 10³–10⁴ at 1.5B. Logit lens at the same positions: 2.7k–15k. The pre-committed choice is visible specifically in the workspace register — this confirms v1's "scale-emergent" conjecture and matches the original paper's Claude-scale findings.

J-lens advantage grows with scale (E1/E2). E2 median best rank 8 (J) vs 19 (LL) (top-10: 55% vs 36%) — the reverse of 1.5B, where LL was better. Consistently, E1 motor convergence is later at 7B (LL next-token agreement only 33% at L26 vs 67% at L25 on 1.5B): representational drift grows with scale, and the Jacobian correction becomes necessary rather than optional.

E4 steering: 21/21 again — third scale/family at 100%.

E6 loading replicates, strengthens, and clarifies: J-lens soul-battery Δ = 0.23–0.27 log-rank units (control Δ ≤ 0.08), behavior 1.00 vs 0.67, r(occupancy, behavior) = −0.95 (1.5B: −0.80; Llama-1B: −0.87). Notably the logit lens's loading contrast washes out at high dilution on 7B while the J-lens's remains — at scale, the workspace register needs the corrective lens.

Honest negatives at 7B: (a) coarse subspace dissociation still absent (proxy basis = 0.06% of variance; ablating it is now essentially harmless — NLL 3.37 vs random 3.66); (b) the rank-1 targeted knockout that killed 3/7 items at 1.5B has zero selective kills at 7B — larger models are more redundant; single directions stop being single points of failure, consistent with the original's use of full J-space cones rather than single vectors.