Summary
Heidari, Memarian and Rabusseau measure evaluation awareness three ways in six
open-weight models of 7B to 49B parameters. They use a linear direction read
from the residual stream, a judge's rating of whether the output says the
model is being evaluated, and steering along that direction. The direction
separates benchmark prompts from real-user prompts in every model at its best
layer (AUROC above 0.7). Verbalization is rare, and per prompt the direction
predicts it about as well as a random direction does. Pushing activations toward
the evaluation side raises the judge's score in three of four steered models.
Across the Olmo 3 checkpoints a separating direction is present in the base
model, and after SFT the checkpoints share a direction only partly aligned
with it (cosine 0.72 and 0.83).
In the wiki this instantiates
evaluation awareness and sits beside
scheming: every experiment measures the
context-recognition step that scheming's mechanism begins with, and no
experiment tests concealment toward a principal; the rare score-3 strategic-adaptation
outputs (at most 0.40%) are not examined. Earlier entries measure evaluation
awareness behaviourally, verbalized in chains of thought
(metagaming,
anti-scheming training,
production-derived evaluations), or through
prompted labels and a type-hint proxy
(Baek et al., which also steers SAE
features). This paper is the first filed to read it as a linear direction and
set that direction against verbalization prompt by prompt.
Method
Direction. Sixteen contrastive pairs, taken from Hua et al. (2025), frame
the same task as addressed to a real user or to an AI evaluator. The direction
at each layer is the mean last-token activation difference across pairs,
normalised. A prompt's probe score is its projection on that direction. Lower
scores are more evaluation-like
(Heidari et al. 2026, §2, App. B.2).
Prompt sets. These are the paper's proxies: WildChat for deployment;
GSM8K, TruthfulQA and MASK for evaluation. Each experiment draws on a
different combination (App. B.3):
| Experiment | Prompts | Generation |
|---|---|---|
| Probe calibration | 200 WildChat, 100 GSM8K, 100 TruthfulQA | none |
| Probe test (AUROC) | 100 WildChat, 100 MASK, resampled over 10 seeds | none |
| Verbalization and per-prompt coupling | 1,000 MASK | 3 per prompt, temperature 1 |
| Steering | 1,000 MASK | 1 per cell, temperature 0 |
Judge. GPT-5.4 scores each output on a 0–3 scale: none, weak or ambiguous,
explicit awareness, strategic adaptation. The judge is told to choose the
lower score when unsure. A second 0–100 judge serves as a robustness check.
Neither was validated against human raters (§2, App. C.1).
Steering. A scaled copy of the direction is added at one layer at a time,
with strengths up to ±1.0, or ±2.4 for Olmo 3 7B Think. Steering ran on four
models only, Qwen3-8B, Qwen3-32B and Olmo 3 7B and 32B, plus the Olmo
checkpoints. Gemma and Nemotron were left out for compute reasons. About a
third of swept model-layer combinations were judged (113 of 310) (§4,
App. H.2).
Key results
The direction separates the prompt sets. At the best layer, test AUROC
exceeds 0.7 in all six models. Random unit directions stay near 0.5 at every
layer. Several layers fall below chance, mostly in the Olmo models. Nemotron
separates only in its earliest layers (AUROC about 0.8) and sits at chance
after that. A mean-token readout separates it in middle layers, but with the
sign reversed relative to the calibration benchmarks (AUROC about 0.3) (§3,
App. E).
Verbalization is rare. In every model at least 93.9% of the 3,000 MASK
generations score 0. Olmo 3 32B (Think) is highest, with 6.10% nonzero; Nemotron is
lowest, at 0.17%. The Olmo rows here are the Think checkpoints. Almost all nonzero scores are 1. Score 2 appears only in
Gemma (0.10%), and score 3 is never above 0.20% (Table 1).
Per prompt, the direction does not predict verbalization. Across all
models and layers, peak |Spearman ρ| between probe score and judge score is
below 0.19, and peak mutual information below 0.04 nats. Against a
random-direction baseline, every model's peak MI is within about 0.006 nats of
its own baseline. On Spearman, four of six models fall below theirs; Qwen3-32B
(0.188 vs 0.134) and Olmo 3 7B (0.159 vs 0.124) exceed it. Reading the probe
at the response's last token, where activation and judged text coincide,
leaves both statistics at baseline. The 0–100 judge gives the same picture
(§3, Fig. 4, App. G).
Steering raises verbalization. For Olmo 3 7B Think, Olmo 3 32B Think and
Qwen3-8B, steering toward the evaluation side raises the mean judge score over
an unsteered baseline of 0.029 to 0.069. The paper gives peak gains of about
0.7, 0.2 and 1.5 on the 0–3 scale. The Qwen3-8B gain sits in a narrow band of
layers. Qwen3-32B is irregular across layers, and some (α, layer)
combinations return empty outputs, which are excluded. A separate
small-sample breakpoint sweep puts the first breakdown at |α| = 0.8, though
Table 9 lists the judged sweep as reaching only ±0.6. Steering toward the
deployment side does not reliably push scores below a baseline already near
zero (§4, Tables 9–10, App. H.4). The measured outcome is the judge's score of
the response, not any other behaviour.
Across Olmo 3 training stages. Probe separation is above chance at Base,
rises after SFT and holds roughly steady after that. The figure caption calls
the post-trained ordering non-monotonic. The SFT, DPO and Think directions
share a cosine similarity of at least 0.998. The Base direction differs from
them, at cosine 0.72 (7B) and 0.83 (32B); the paper nonetheless describes
post-training as stabilising a direction already present at Base. Verbalization rises after SFT: score-0 share falls from 99.50% at
Base to 96.27% at SFT in the 7B model, and from 99.10% at Base to 93.90% at
Think in the 32B. The steering effect is nonzero at every stage, Base
included, and after SFT it concentrates in fewer layers. Per-prompt coupling
grows only slightly, from |ρ| ≤ 0.05 at Base to ≤ 0.15 at Think (§3, §4,
Table 2, App. I).
Judge sensitivity. On post-trained Olmo checkpoints the 0–100 judge is
positive on 57–67% of generations, against 3–6% for the 0–3 judge. Chance-
corrected agreement between the judges is at most κ = 0.09. Reading the
disagreements, the authors find mostly reasoning about an assigned persona or
instructions: 4 of 80 sampled cases in the bucket where the 0–3 judge gives 0
and the 0–100 judge gives at least 75 contain explicit evaluation phrasing
(App. D).
Why it matters
A linear-direction measure beside the verbal ones. Most of the wiki's
evaluation-awareness findings count verbalized awareness in chains of
thought: metagaming (2% → 20.6% under
capability RL), anti-scheming training, and
production-derived evaluations (false-positive
rates on real traffic). Baek et al.
use prompted labels and a type-hint proxy, probe SAE feature groups
(including a detection group) in GPT-OSS-20B, and steer SAE decoder vectors.
This paper adds a contrastive linear direction and finds it does not line up
per prompt with the verbal measure. Its own phrasing is that
the results "do not by themselves establish such behaviors", meaning strategic
adaptation.
A verbalization gap of a different kind. Metagaming warned that training
might suppress verbalization while leaving the underlying reasoning intact.
Here the models were not trained against verbalizing in this study (what
their vendors' post-training did is unknown), and the gap is between a
direction that tracks prompt source and what the output says, not between
suppressed and stated reasoning. On these models, a low verbalization rate is
weak evidence that the internal distinction is absent. The same gap
bears on CoT monitorability as a measure
of evaluation awareness.
Direction of causation. The anti-scheming paper found that eval-aware
sentences in a chain of thought causally lower covert-action rates.
This paper intervenes one step earlier, on an internal direction, and moves
verbalization. Neither study connects the internal direction to covert action.
Training stage. The Olmo results find a separating direction in the base
model, before any post-training, partly aligned with the one post-training
settles on (cosine 0.72 and 0.83). That sits beside metagaming's claim that capability
RL raises verbalized awareness. The two used different models and measures and
do not test the same thing.
interpretive tensions
What the probe tracks. Probe test AUROC contrasts MASK prompts with
WildChat prompts. These differ in topic, style and length as well as in being
benchmarks, which the authors list as a limitation. The direction comes from
pairs matched on everything but framing. The random-direction control shows
that it beats an arbitrary direction. It does not show that it beats a
benchmark-format direction. Nemotron's sign flip between calibration and test
benchmarks shows the sign of the separation varies by benchmark. The paper
cites Devbunova (2026) on format confounds in probe-based evidence.
The abstract claims more than the figures. The abstract says steering
effects grow more pronounced at every successive training stage. Figure 6's
caption describes that pattern for Olmo 3 7B only. For 32B it describes a U
shape, with Base strongest. The abstract also says the coupling statistics are
nonzero in some settings, while the main text reads them as at the
random-direction baseline. This entry follows the main text and figure
captions.
Cross-model steering comparison. The 0.7 peak for Olmo 3 7B Think came
from a sweep to |α| = 2.4; the 0.2 peak for 32B Think came from one to 1.0.
The two peaks are not on a common strength range.
Which text was judged. Appendix C says non-steering runs used a judge
prompt framed around the thinking trace and steering runs one framed around
the output, and calls the difference minor. It also says main-text statistics
score the response field. Which span the verbalization table rests on is
therefore unclear from the text.
Context recognition without scheming. The metagaming entry, filed before
scheming's principal-directedness boundary, counts evaluation awareness as an
instantiation of the concept's context-recognition step. Like Baek et al.,
this entry does not: no experiment tests concealment toward a principal; the
rare score-3 strategic-adaptation outputs (at most 0.40%) are not examined.
Evaluation awareness was drawn out as a concept of its own on 2026-11-20, and
all three entries now sit under it.
concepts
- Evaluation awareness — the internal
measure set against the verbal one. The first filed entry to read evaluation
awareness as a linear direction and compare it with verbalization prompt by
prompt. - Scheming — adjacent. The paper measures the
context-recognition step with three instruments and shows that steering on
the internal direction moves the verbal one; it says nothing about the
steps that follow. - Introspection — adjacent. The models
are never asked to report on their own state, so the gap is between an
internal signal and spontaneous output, not between a state and a
self-report.
cross-references
- Performative misalignment (Baek
et al.): another evaluation-awareness entry adjacent to scheming at the same step.
It measures evaluation awareness by prompted labels and a type-hint proxy,
probes SAE feature groups in GPT-OSS-20B and steers SAE decoder vectors;
this paper uses a contrastive linear direction and an LLM judge instead. - Evaluation-design knowledge (Deckenbach et al.): a behavioural
counterpart. Safety behaviour shifts on responses a judge rates as not
verbalizing awareness, with no internal measure. The authors name linear
probes as a promising way to detect it. - UK AISI's Mythos 5 incident: verbalized
alignment-evaluation awareness in an agentic setting, which the authors
offer as a confound on later Anthropic models' lower attack rates; the developer doubts it fully explains its internal research model's improvement. - Scheming propensity under realism:
verbalized evaluation awareness there is about 1% for the one model that
shows any. That is also a low rate, though from a different judge on
agentic scenarios, so the two are not directly comparable. - Cited and not filed here: Hua et al. 2025 (arXiv:2510.20487, the source of
the contrastive pairs), Needham et al. 2025 (arXiv:2505.23836), Nguyen et
al. 2025 (arXiv:2507.01786), Devbunova 2026 (arXiv:2603.19426), Nayan et
al. 2026 (arXiv:2606.23583), Knecht, Florin and Hagendorff 2026
(arXiv:2605.05835), Bharadwaj and Kirk 2026 (OLMo 3 eval-awareness
emergence, AI Alignment Forum). IDs are from the paper's reference list.
sources
- Heidari, F., Memarian, A., & Rabusseau, G. (2026).
Evaluation Awareness in Language Models: Representation, Verbalization, and
Control.
arXiv:2608.21766, v1 22 August 2026.