Probe Guidance: the weak model you need for guidance is already inside the strong one
Date: 2026-09-20
Topic: llms-foundation-models
Source: X home feed (@shumpeiMaxwell)
Link: arXiv:2609.19356
Authors: Rohit Dilip, Tianrong Chen, Yuyang Wang, David Van Valen, Josh Susskind, Miguel Angel Bautista (Apple; Caltech)
Enrichment: alphaxiv overview available
Raw: raw/twitter/feed/2026-09-20-morning-ranked.json
TL;DR
Diffusion language models generate a sequence by iteratively refining a noisy representation rather than emitting tokens left to right. They inherit guidance techniques from image diffusion, where you improve sample quality by extrapolating away from a weaker prediction. The standard version, autoguidance, needs an actual weak model (an earlier checkpoint, or a degraded copy), and running it roughly doubles inference compute. Classifier-free guidance, the other standard option, needs a class condition that language generation usually does not have.
Probe Guidance removes the second model. It trains a small multilayer perceptron on the strong model's own intermediate hidden states to act as the weak predictor, then guides using the strong-minus-weak difference. The backbone stays frozen. The probe reads activations the forward pass already computed, so there is no second full model evaluation. The result is guidance at roughly the cost of no guidance.
flowchart LR
X[Noisy sequence<br/>at step t] --> BB[Frozen diffusion LM<br/>backbone]
BB --> H[Intermediate<br/>hidden states]
BB --> SP[Strong prediction]
H --> PR[Tiny MLP probe<br/>trained separately]
PR --> WP[Weak prediction]
SP --> G[Guided output =<br/>strong + w x strong-minus-weak]
WP --> G
G --> NX[Next refinement step]
ALT[Autoguidance baseline:<br/>second full model pass] -.->|~2x compute| WP
classDef input fill:#dbeafe,stroke:#3b82f6,color:#1e3a8a
classDef decision fill:#fef3c7,stroke:#f59e0b,color:#78350f
classDef output fill:#d1fae5,stroke:#10b981,color:#065f46
classDef warn fill:#fee2e2,stroke:#ef4444,color:#7f1d1d
classDef aux fill:#e0e7ff,stroke:#6366f1,color:#312e81
class X input
class BB,H aux
class PR,SP,WP decision
class G,NX output
class ALT warn
Why the mechanism is more interesting than the benchmark
The paper's own positioning is efficiency: get autoguidance's benefit without autoguidance's second forward pass. That is a clean inference-cost win and it is real. But the reason this matters to this wiki is that it is the third independent result in three days arguing that a model's intermediate hidden states already contain a signal the field has been paying a separate model to compute.
- Gavel (09-16) claims the skill routing signal is already present in an agent model's mid-layer hidden states and should be read out with a lightweight head rather than derived from a separate router model.
- C2C (09-18) projects one model's KV cache directly into another's with a learned per-layer gate, on the premise that forcing model-internal state through human vocabulary destroys information, reporting a 2.5x speedup over text handoff.
- Probe Guidance trains a tiny MLP on intermediate activations to serve as the weak model in a guidance pair, replacing a full second evaluation.
Three different subfields (agent routing, multi-agent communication, diffusion sampling), three different target quantities (which skill, what the other model knows, what a worse model would have predicted), and the same architectural move: stop re-deriving it, read it off the residual stream. The 09-18 routing entry called C2C "a point on the board for reading internals rather than re-deriving them through a separate model." This is the second point, and the threshold for declaring a pattern on this wiki is three. The pattern is now established: the intermediate representation is an underexploited free resource, and the field has been buying a second model to recompute what it already has.
The contrast with the day's other dominant story is sharp and worth stating. The Jev ecosystem (09-20) is 160 projects betting that a separate, purpose-built, external model should make the cheap decisions. Gavel, C2C and Probe Guidance are three results betting the signal is internal and already paid for. These are opposite answers to the same question and neither camp has run the other's experiment.
Method details worth keeping
Autoguidance works only when the strong and weak models have related dynamics, which is why degradation tricks like isotropic dropout can produce a weak model that guides in an unhelpful direction. Probe Guidance gets that relatedness structurally rather than by luck: the weak prediction is a function of the strong model's own activations, so the two prediction paths are coupled by construction. The authors flag this as a design advantage and it is the most defensible part of the argument.
Because the backbone is frozen and the probe is separate, the method is applicable to released checkpoints and easy to ablate. That is a deployment property, not just a research convenience.
The experiments target continuous diffusion language models, covering both latent-space and data-space formulations. This is a narrower setting than "language models" and should not be read across to autoregressive systems without evidence.
Gaps
The obvious one is scope: continuous diffusion language models are a small and fast-moving corner, and it is unclear whether the probe trick survives in discrete-diffusion or autoregressive settings where there is no analogous guidance step. The probe must be trained, so the "free" claim is about inference cost and not total cost, and the abstract-level material does not state how much data the probe needs. No head-to-head against the internal-signal results in adjacent areas exists, which is exactly the comparison that would turn three isolated findings into a usable principle.