jlens — read a model's internal draft

rapid-mlx jlens "<prompt>" is a read-only interpretability command. It reads what a model is disposed to say at every layer using the Jacobian lens (J-lens), then reports where the answer crystallizes, how much depth is left unused, and how far ahead of the plain logit lens it can see — locally, on your quantized MLX models, with no server and no telemetry.

What it does

The logit lens decodes an intermediate residual-stream vector directly with the unembedding — but at early and middle layers that often reads as noise, because a layer's activation also matters for future tokens, not just the next one. The Jacobian lens corrects for that: for each layer l it estimates the corpus-averaged transport

J_l = E[ ∂h_final / ∂h_l ]

and decodes unembed(J_l · h) with the model's own unembedding. Because J_l accounts for an activation's influence on the final layer, J-lens surfaces intermediate reasoning steps and concept clusters that the logit lens misses — and locks onto the answer several layers earlier.

Quick start

$ rapid-mlx jlens "The Eiffel Tower is located in the country of"

No model flag runs the default qwen3-1.7b. First use pulls the weights (a few GB) after the same confirmation prompt the other weight-loading commands use; a small model takes roughly 6–8 seconds including load. Output:

rapid-mlx jlens · qwen3-1.7b · 'The Eiffel Tower is located in the country of'
  model continues → 'France, and it is the'
  [jvp]  28 layers

internal trajectory  (J-lens: what the model is disposed to say, by depth)
  early  L 0-6   ·  深度  ipples  一步步  回  以上  norm
  mid    L 7-13  ·  Land  土地  Green  测量  Germany  (country
  deep   L14-20  ·  Greece  Albania  Spain  США  Appalach  República
  final  L21-27  ·  France  Spain  Germany  法国  Belgium  French

answer 'France':
  crystallizes at  L12/28  (43% depth)  →  early-exit headroom ~57%
  J-lens lead over logit lens:  +10 layers  (logit lens locks at L22)

workspace signal  0.83  ########..   (experimental heuristic)
  mid-layer concept density 95%  ·  J-lens lead 10L

Reading the output

• internal trajectory — the top J-lens tokens grouped by depth band. Early layers read as noise; by the middle band generic scaffolding appears (country, Land), the deep band narrows to a European-country cluster, and the final band commits to France. That mid-band cluster is exactly what the plain logit lens misses.
• crystallizes at Lk — the layer where the answer token first takes and holds the lead. early-exit headroom is the fraction of depth after that point — a rough signal of how much of the network this prompt did not need.
• J-lens lead over logit lens — how many layers earlier J-lens reads the answer than the plain logit lens. This is the whole point: earlier, cleaner signal.
• workspace signal — an experimental heuristic (0–1) inspired by the paper's global-workspace framing. It partly reflects answer confidence and is not a validated capability metric — read it as a hint, not a measurement. The line beneath it recaps the inputs it leans on: mid-layer concept density (how much of the middle band carried nameable concepts) and the J-lens lead from above.

Model size changes what you see

The default qwen3-1.7b is sized for a fast first run, not for showing off the method. On a model that small the intermediate bands are mostly noise and the J-lens lead is often +0 — the answer only crystallizes near the output. The global-workspace structure the method is built to surface is a property of larger models; it starts to show clearly around 8B.

The same prompt on Llama-3.1-8B — watch the deep band represent the category ('city') before it commits to the instance ('Tokyo'), a genuine intermediate-reasoning trace, with a positive J-lens lead:

rapid-mlx jlens · Meta-Llama-3.1-8B-Instruct-4bit · 'The capital of Japan is'
  model continues → 'Tokyo. The capital of Japan'

internal trajectory  (J-lens: what the model is disposed to say, by depth)
  early  L 0-7   ·  facts  columns  UGIN  relation  truths  persons
  mid    L 8-15  ·  worldwide  Lv  USA  FOREIGN  International  Global
  deep   L16-23  ·  city  cities  metropolitan  Tokyo  Taipei  Beijing
  final  L24-31  ·  Tokyo  Seoul  Kyoto  東京  Osaka  Tok

answer 'Tokyo':
  crystallizes at  L18/32  (56% depth)  →  early-exit headroom ~44%
  J-lens lead over logit lens:  +4 layers  (logit lens locks at L22)

workspace signal  0.42  ####......   (experimental heuristic)
  mid-layer concept density 58%  ·  J-lens lead 4L

Two things hold at every size: tight lexical lookups (a chemical symbol, say) crystallize late with little intermediate structure, while multi-hop and conceptual prompts show the most; and the earliest layers read as noise regardless — the workspace lives in the middle and deep layers. For real interpretability work, reach for a model of 8B or larger via -m; keep the 1.7B default for a quick look.

Options

How it stays cheap — and works on quantized models

• No Jacobian is ever materialized. J_l · h is computed with a Jacobian–vector product (forward-mode autodiff), so the full d×d matrix never exists in memory. Cost is O(corpus_size) forward passes per layer.
• It differentiates through 4-bit / 8-bit weights. The JVP is taken with respect to activations — the quantized weights are constants along the way — so jlens runs directly on your cached MLX models with no dequantization step. A finite-difference path is used as a fallback where forward-mode autodiff is unavailable.

Supported models

Standard dense decoder transformers — Qwen3, Llama, Phi, and similar architectures (including decoder blocks that require a cache argument, e.g. StableLM). Architecture detection is automatic.

Linear-attention / hybrid models (e.g. Qwen3.5) and the vision towers of VLMs are detected and reported as unsupported rather than crashing — you get a clear message, not a stack trace.

Limitations

• The workspace signal is an experimental heuristic, not a validated capability metric; it partly reflects answer confidence. A rigorous version would calibrate against a J-space ablation on multi-step tasks.
• The phenomenon is scale-dependent. On small models (≈1–2B) the intermediate trajectory is largely noise and the J-lens lead collapses toward zero; meaningful workspace structure emerges around 8B and grows with size — see Model size changes what you see.
• Coverage is dense decoders today; linear-attention / hybrid architectures and VLM vision towers are out of scope and flagged as such.

Where to next

Read the method in Anthropic's Verbalizable Representations and Global Workspace in Language Models. For the rest of the CLI — serve, pull, launch — see the engine README and the HTTP API surface.