Every number, and what it rests on

Each entry gives the model, dataset, sample size, method and source. Paper values were checked against the arXiv PDFs. Read the caveats alongside the numbers: most results come from one model or a small sample.

Where is it computed?

L17–35

Paper

Four preference-localization methods converge on one subgraph.

Subgraph-level result on a single financial task. A separate classification pipeline surfaces an MLP cluster at L8–L11.

Model
Qwen3-4B-Instruct-2507
Dataset
500 explicit and 500 implicit pairs, plus 4,588 parametric prompts
Method
Probing, contrastive attribution, parametric attribution and parametric patching
Source
Rios-Sialer et al., Section 5.1, Figure 1

L24 and L21 attention; L35 and L31 MLPs

Paper

Activation patching ranks four components above the rest.

Finer-grained circuit tracing has not been done.

Model
Qwen3-4B-Instruct-2507
Dataset
Parametric prompt pairs
n
71
Method
Parametric activation patching, ranked by mean causal effect
Source
Rios-Sialer et al., Section 5.1, Figure 3

What form does the horizon take?

Seconds to centuries

Paper

Time horizons form ordered clusters on a curved manifold.

Shown with PCA projections. The main text reports no quantitative separation metric.

Model
Qwen3-4B-Instruct-2507
Dataset
4,588 investment choice prompts with optional horizons
n
4,588
Method
PCA inside the localized subgraph
Source
Rios-Sialer et al., Section 5.2, Figures 1 and 5

When does it become a preference signal?

User-to-assistant boundary

Paper

Attention collapses the horizon representation into a binary preference signal at the turn boundary.

Separation is shown visually. No-horizon samples align with the manifold by the assistant token.

Model
Qwen3-4B-Instruct-2507
Dataset
4,588 investment choice prompts with optional horizons
n
4,588
Method
Layer-31 PCA at four turn-transition tokens
Source
Rios-Sialer et al., Section 5.2, Figures 4 and 5

Does a stated deadline reach the output?

47–53%

Paper

With a deadline only the sooner option can meet, outputs still select the undeliverable long-term option about half the time.

Single financial task.

Model
Qwen3-4B-Instruct-2507
Dataset
960 investment prompts, 1–5 year deadline zone
n
960
Method
Investment-coherence questionnaire
Source
Rios-Sialer et al., Section 5.2, Figure 6, Appendix P

95–100%

Paper

Among 30 models on the same instrument, only frontier API models reach high coherence.

The paper reports that smaller Claude variants reach this through a binary heuristic that collapses at longer horizons.

Model
Claude family, Gemini 2.5 Pro, GPT-5.4, GPT-5.4 Mini, o3
Dataset
960 investment prompts
n
960
Method
Investment-coherence questionnaire
Source
Rios-Sialer et al., Section 5.2, Figure 6

3–8×

Paper

Fitted discount rates sit well below human controls.

The paper reports that these preferences are unstable across contexts.

Model
Qwen3-4B and Qwen3-8B
Dataset
Kirby MCQ-27 under 8 conditions
Method
Fitted discount rate k: below 0.005 for the models, about 0.013 for human controls
Source
Rios-Sialer et al., Section 5.2, Appendix O

Does changing the signal change the output?

59.4 points

Paper

Steering at L44 widens the gap in long-term outputs between the strongest negative and positive settings.

A matched random direction moves choices by at most 5.6 points. |α| = 128 is 36–63% of the unsteered residual norm, a deliberately large perturbation. No prompting baseline is reported.

Model
Qwen3-32B
Dataset
Held-out explicit binary-choice prompts
n
100
Method
CAA at L44, greedy decoding: 33.7% at −128, 76.8% at 0, 93.1% at +128
Source
Mráz & Shenk, Section 2, Appendix A Table 6, Table 7

40.6%, 74.3%, 91.4%

Paper

Sampled decoding preserves the steering pattern.

Two-sigma standard errors are between 1.8 and 1.9 points.

Model
Qwen3-32B
Dataset
Held-out explicit binary-choice prompts
n
100
Method
CAA at L44, temperature 0.8, five samples per prompt, at −128, 0 and +128
Source
Mráz & Shenk, Appendix D, Table 7

36.5 points

Paper

The same direction moves outputs on implicit prompts by less.

A matched random direction moves choices by at most 5.5 points.

Model
Qwen3-32B
Dataset
Implicit binary-choice prompts
n
300
Method
CAA at L44, greedy decoding
Source
Mráz & Shenk, Appendix A, Table 6

3.4× odds

Paper

Steering shifts forced-choice scores in both directions.

The abstract calls this suggestive evidence. L21 reaches the same score, and output quality degrades at |α| = 60.

Model
Qwen3-4B-Instruct-2507
Dataset
Held-out forced-choice prompts
n
20
Method
CAA at L22, α = 50: score 0.17 to 1.39, a lift of +1.22
Source
Rios-Sialer et al., Section 5.3, Table 1, Appendix D

Up to +2.7 and −5.7

Paper

Open-ended outputs shift in the steered direction on an LLM-judged scale.

The judge is not validated against human ratings. The paper treats these scores as a proxy for direction, not a calibrated measure.

Model
Qwen3-4B-Instruct-2507
Dataset
Open-ended neutral prompts, scored from −10 to +10
n
13
Method
CAA at L19–23 and L26, α up to 50, scored by Claude Sonnet 4.6
Source
Rios-Sialer et al., Appendix S.2, Table S.3

Direction only

Paper

Moderate long-term steering raises a planning-related capability metric, and +128 lowers it.

The effect is moderate and non-monotonic. The text gives no significance test, and the authors say it is not evidence that steering generally solves planning.

Model
Qwen3-32B
Dataset
TravelPlanner validation queries
n
180
Method
CAA at L40, Commonsense Constraint Micro Pass Rate
Source
Mráz & Shenk, Section 4, Figure 5

Is it really about time?

99.2%

Paper

A linear probe separates short- from long-term labels on prompts checked for explicit temporal keywords.

Shuffled labels score about 50%, which rules out chance but not surface cues: no surface baseline yet. The same probe scores about 80% at layer 0.

Model
Qwen3-4B-Instruct-2507
Dataset
500 implicit pairs, 80/20 pair-aware split
n
500
Method
Logistic regression at L26, test accuracy
Source
Rios-Sialer et al., Appendix G.1, Table G.1

Above 95%

Paper

Probes trained on implicit pairs carry over to explicit pairs.

Transfer shows the two sets share a readable contrast, not what that contrast is: no surface baseline yet.

Model
Qwen3-4B-Instruct-2507
Dataset
Trained on implicit pairs, evaluated on explicit pairs
Method
Zero-shot transfer at L26
Source
Rios-Sialer et al., Appendix G.4

99.0%

Paper

Probes trained on explicit questions read a held-out explicit set.

The authors treat this as evidence the contrast is linearly accessible, not that it isolates temporal horizon: no surface baseline yet.

Model
Qwen3-32B
Dataset
Explicit questions, 400 train and 100 test
n
100
Method
Best probe families at L28 and L32; the steering direction reaches 96.0% at L24
Source
Mráz & Shenk, Section 1.2, Table 2

77.0–83.5%

Paper

The same probes score lower on implicit questions.

The authors say implicit examples do not fully isolate horizon: no surface baseline yet.

Model
Qwen3-32B
Dataset
Implicit questions, full set
n
300
Method
Explicit-trained probes, best layer per family between L24 and L32
Source
Mráz & Shenk, Section 1.2, Table 2

About 95%

Paper

An unrelated variable, cumulative reliability of a reasoning chain, is also decodable in the same subgraph.

The direction is linearly orthogonal to temporal preference but on the same curved manifold, so the region encodes more than time.

Model
Qwen3-4B-Instruct-2507
Dataset
Error-injected multi-hop math reasoning steps
n
3,550
Method
Logistic regression in shared PCA-50 space, plateau across L19–31
Source
Rios-Sialer et al., Section 5.2, Appendix R

Open

Paper

The steering direction may reflect a broader short-term/long-term register.

Continuations can also vary in abstraction, urgency, strategic scope and proactive-versus-reactive framing.

Model
Qwen3-32B
Dataset
Explicit answer continuations
n
800
Method
Stated limitation on the training contrast
Source
Mráz & Shenk, Section 5.2

How far does this generalize?

8.34, 2.67, 1.12

Paper

A direction trained on binary answers moves the indifference multiplier on a monetary task it was not trained on.

A random direction gives a negative-to-positive ratio of 0.39–1.69, against 4.55–14.46 for steering. L40 is an illustrative slice, and the 3650-day value is capped at the largest tested multiplier.

Model
Qwen3-32B
Dataset
Money-now versus money-later prompts, 768 generations per condition
n
768
Method
CAA at L40, greedy decoding, at −128, 0 and +128
Source
Mráz & Shenk, Section 3, Table 5

Varies by family

Project report

Probe signal differs across model families.

Reported as a qualitative comparison. Cross-model causal patching has not been done.

Model
Qwen3, Llama, Phi-3-mini
Dataset
Fellow-specific contrastive datasets
Method
Linear probes compared across models
Source
SPAR final report, Section 3.1

84%

Repository result

An explicit-trained probe reads implicit prompts in a much smaller model.

One small model and a small set: no surface baseline yet.

Model
GPT-2
Dataset
Trained on explicit pairs, evaluated on 50 implicit pairs
n
100
Method
Logistic regression at L10
Source
GPT-2 probe validation, implicit_results

Numbers this site no longer features

These earlier GPT-2 numbers remain here so readers can trace what changed. An earlier steering-correlation figure is withdrawn altogether: its plot is missing and its verification never resolved.

92.5%

Repository result

Earlier headline number for GPT-2.

A training-set score, so not evidence of generalization: no surface baseline yet.

Model
GPT-2
Dataset
Explicit pairs
n
400
Method
Cross-validation on the training set at L8
Source
GPT-2 result files, full_train_results.csv

84%

Repository result

Earlier held-out number for GPT-2.

The file records cross-validation over its own 100 samples, not a train-to-test result: no surface baseline yet.

Model
GPT-2
Dataset
Test file
n
100
Method
Cross-validated accuracy recorded within the test file at L6
Source
GPT-2 result files, full_test_results.csv

100%

Unverified

Earlier claim of semantic encoding after keyword removal.

Accuracy rises from 74–77% in the test file to 100% after keywords are removed, which needs explaining. The verification script's quick mode only prints stored values.

Model
GPT-2
Dataset
Keyword-ablated control file
n
100
Method
Cross-validated accuracy recorded at L10–11
Source
GPT-2 result files, control_ablated_results.csv

~60%

Repository result

Earlier steering consistency figure for GPT-2.

Specific to GPT-2, and superseded by the controlled Qwen3 sweeps above.

Model
GPT-2
Dataset
Temporal steering prompts
Method
Listed as a known caveat in the results summary
Source
GPT-2 results summary, Known Caveats