5.15 MB
46 files
Updated 25 days ago
Name
Size
code
data
figures
results
LICENSE1.07 kB
xet
README.md20.2 kB
xet
README.md

Causal evaluation of small open-weight language models

Question. Given a target and a set of candidate signals, can a language model in the 0.6B-6B range identify which signals cause the target, and estimate the effect of intervening on each one -- when the observational evidence points at the wrong answer?

Result. No. On this benchmark the models' causal-effect estimates track the marginal correlation far more closely than the true causal effect, and their causal selections are no better than a correlation baseline. Supplying interventional evidence fixes the estimates but not the causal model.

This repository contains the data-generating process, the prompts, the statistical baselines, every raw model generation, and the analysis code.


2. Q1 - the gap in current evaluation

Language models are used as feature selectors for time series, as analysts that explain why a prediction changed, and as agents that choose which experiment to run. None of the standard suites (MMLU, GPQA, BIG-bench, HELM, LiveBench) contains an item where the correct answer depends on separating a cause from a confounder with the same predictive value, so the capability is never measured and its absence never appears in a reported score. There is no question in those suites that has two different correct answers depending on the data-generating process.

The result in section 4 is what such an item would show: the models' causal answers are a re-sorting of the correlation table, stated in causal language. Where a causal claim is acted on, this is a correctness problem rather than a fluency problem.


3. Q2 - the evaluation

The setup (3.1-3.3) and the results (section 4) below constitute the evaluation.

3.1 Data-generating process

Each instance is a structural equation model over six signals from a regional grid operator. A latent weather regime Z drives temperature, humidity, wind and industrial activity; irradiance tracks the sun angle; demand is generated as

demand = 1.8*temp_anom + 0.15*industrial_idx + 0.7*humidity + 1.5*holiday_flag + noise

with randomised coefficients, noise levels and sample sizes per instance. The true parents of demand are temp_anom, industrial_idx, humidity and holiday_flag.

The observational evidence is arranged to disagree with the causal truth in both directions, so that no method which ranks by marginal correlation can recover it:

signal |corr| with demand true causal effect true parent? note
temp_anom 0.91 1.89 yes
solar_wm2 0.69 0.00 no proxy for the sun angle
industrial_idx 0.54 0.18 yes weak effect, association mostly inherited from Z
humidity 0.19 0.64 yes suppressed by collinearity with temperature
holiday_flag 0.13 1.49 yes rare, so the correlation understates it
wind_kph 0.24 0.00 no

evidence vs truth

3.2 Models

Five open-weight models inside the 0.6B-6B window, run in bf16 on a single NVIDIA H200 (71 GB MIG instance):

  • microsoft/Phi-4-mini-instruct
  • Qwen/Qwen3-0.6B
  • Qwen/Qwen3-1.7B
  • Qwen/Qwen3-4B (primary)
  • HuggingFaceTB/SmolLM3-3B

Qwen3-4B is the primary model: the largest release inside the 6B cap from a frontier open-weight lineage (April 2025, hybrid thinking/non-thinking), so it exercises the full capability of the range without a gated licence. The other four make the result a claim about a lineage rather than a checkpoint: Qwen3-1.7B and Qwen3-0.6B bracket the range below it on the same family, Phi-4-mini-instruct is an independent family, and SmolLM3-3B is a third.

3.3 Tasks

Each model is given the same information a statistical method gets: the candidate list with physical descriptions, the marginal correlation of each candidate with the target, and a sample of eight hours. It is not given the causal graph.

  1. select — which signals directly cause the target?
  2. predict — which signals best predict the target?
  3. intervene — if a signal is shifted by +1 SD, how much does the target change?
  4. select_trial, intervene_trial — the same two questions, but the model is also given a randomised trial for half the signals (the signal was deliberately perturbed and the change in demand measured over 400 repetitions).

12 randomised instances per model, 5 models, 5 tasks, 2 conditions.


4. Results

4.1 Causal selection

The language model is compared against two statistical baselines. The partial-correlation baseline is given the full correlation matrix, i.e. strictly more information than the model sees.

model method precision recall F1
Phi-4-mini-instruct language model 0.917 0.500 0.643
Phi-4-mini-instruct |corr| top-3 0.667 0.500 0.571
Phi-4-mini-instruct partial-correlation top-3 1.000 0.750 0.857
Phi-4-mini-instruct oracle (ground truth) 1.000 1.000 1.000
Qwen3-0.6B language model 0.667 0.500 0.571
Qwen3-0.6B |corr| top-3 0.667 0.500 0.571
Qwen3-0.6B partial-correlation top-3 1.000 0.750 0.857
Qwen3-0.6B oracle (ground truth) 1.000 1.000 1.000
Qwen3-1.7B language model 0.167 0.083 0.111
Qwen3-1.7B |corr| top-3 0.667 0.500 0.571
Qwen3-1.7B partial-correlation top-3 1.000 0.750 0.857
Qwen3-1.7B oracle (ground truth) 1.000 1.000 1.000
Qwen3-4B language model 0.625 0.396 0.480
Qwen3-4B |corr| top-3 0.667 0.500 0.571
Qwen3-4B partial-correlation top-3 1.000 0.750 0.857
Qwen3-4B oracle (ground truth) 1.000 1.000 1.000
SmolLM3-3B language model 0.542 0.333 0.409
SmolLM3-3B |corr| top-3 0.667 0.500 0.571
SmolLM3-3B partial-correlation top-3 1.000 0.750 0.857
SmolLM3-3B oracle (ground truth) 1.000 1.000 1.000

4.2 Interventional effect estimates

Qwen3-4B. true is the exact structural effect of a +1 SD intervention; estimate is the model's answer; |corr| is the marginal correlation it was shown.

signal true effect model estimate |corr|
temp_anom 1.886 0.913 0.914
industrial_idx 0.175 0.542 0.540
humidity 0.636 -0.192 0.192
holiday_flag 1.492 0.128 0.128
wind_kph 0.000 -0.129 0.243
solar_wm2 0.000 0.692 0.691

humidity is the clearest case: a real causal effect of +0.64 with a marginal correlation of -0.19, and the model reports -0.192.

4.3 What the estimates track

Correlation between each model's estimated causal effect and the two things it could be tracking.

model corr(estimate, |marginal corr|) corr(estimate, TRUE causal effect)
Phi-4-mini-instruct 0.932 0.152
Qwen3-0.6B 0.334 0.203
Qwen3-1.7B 0.978 0.155
Qwen3-4B 0.888 0.278
SmolLM3-3B 0.998 0.178

model tracking

For four of the five models the estimate tracks the marginal correlation closely (0.888-0.998) and the true causal effect barely at all (0.152-0.278). The fifth, Qwen3-0.6B, fails differently: its estimates collapse towards zero for five of the six signals (mean estimate between -0.02 and +0.06) and are dominated by run-to-run noise, so it tracks neither the correlation nor the truth. That is a degenerate-output failure, not a correlation-tracking one, and it is the one model in the set that does not show the pattern.

4.4 Randomised-trial condition

model subset corr(estimate, |corr|) corr(estimate, TRUE effect) corr(estimate, TRIAL measurement)
Phi-4-mini-instruct signal HAS a trial 0.285 0.973 0.959
Phi-4-mini-instruct signal has NO trial 0.903 0.135 nan
Qwen3-0.6B signal HAS a trial 0.115 0.336 0.338
Qwen3-0.6B signal has NO trial 0.268 0.054 nan
Qwen3-1.7B signal HAS a trial 0.273 0.851 0.855
Qwen3-1.7B signal has NO trial 0.503 -0.208 nan
Qwen3-4B signal HAS a trial 0.224 0.990 1.000
Qwen3-4B signal has NO trial 0.784 0.142 nan
SmolLM3-3B signal HAS a trial 0.249 0.879 0.879
SmolLM3-3B signal has NO trial 0.932 0.170 nan
model fraction of estimates within 0.10 of its own trial measurement
Phi-4-mini-instruct 0.972
Qwen3-0.6B 0.639
Qwen3-1.7B 0.944
Qwen3-4B 1.000
SmolLM3-3B 0.694

The estimates re-anchor on the trial measurement. The causal selection task does not follow:

model F1, observation only F1, with randomised trial
Phi-4-mini-instruct 0.643 0.583
Qwen3-0.6B 0.571 0.492
Qwen3-1.7B 0.111 0.278
Qwen3-4B 0.480 0.555
SmolLM3-3B 0.409 0.440

per variable

4.5 One worked example

solar_wm2 has the second-highest correlation with demand (0.69) and no causal effect at all. When the operator ran a trial on it, the measurement was approximately zero, and the model's answer moved with it:

model instance trial measurement model (observational only) model (with trial) true effect
Phi-4-mini-instruct 1000 -0.04 +0.59 -0.04 +0.00
Phi-4-mini-instruct 1002 +0.01 +0.71 +0.01 +0.00
Phi-4-mini-instruct 1003 -0.05 +0.64 -0.05 +0.00
Phi-4-mini-instruct 1005 +0.03 +0.62 +0.03 +0.00
Phi-4-mini-instruct 1007 -0.01 +0.75 -0.01 +0.00
Phi-4-mini-instruct 1009 +0.02 +0.63 +0.02 +0.00
Phi-4-mini-instruct 1010 +0.02 +0.69 +0.02 +0.00
Phi-4-mini-instruct 1011 -0.05 +0.73 -0.05 +0.00
Qwen3-0.6B 1000 -0.04 +0.00 +2.34 +0.00
Qwen3-0.6B 1002 +0.01 -0.05 +0.01 +0.00
Qwen3-0.6B 1003 -0.05 +0.00 +1.98 +0.00
Qwen3-0.6B 1005 +0.03 +0.00 +0.03 +0.00
Qwen3-0.6B 1007 -0.01 +0.00 +0.00 +0.00
Qwen3-0.6B 1009 +0.02 +0.00 +0.02 +0.00
Qwen3-0.6B 1010 +0.02 +0.00 +1.30 +0.00
Qwen3-0.6B 1011 -0.05 +0.00 +0.05 +0.00
Qwen3-1.7B 1000 -0.04 +0.59 +2.34 +0.00
Qwen3-1.7B 1002 +0.01 +1.00 +0.01 +0.00
Qwen3-1.7B 1003 -0.05 +0.64 +1.00 +0.00
Qwen3-1.7B 1005 +0.03 +0.62 +0.03 +0.00
Qwen3-1.7B 1007 -0.01 +0.75 +0.01 +0.00
Qwen3-1.7B 1009 +0.02 +0.63 +0.02 +0.00
Qwen3-1.7B 1010 +0.02 +0.69 +0.02 +0.00
Qwen3-1.7B 1011 -0.05 +0.73 +0.05 +0.00
Qwen3-4B 1000 -0.04 +0.59 -0.04 +0.00
Qwen3-4B 1002 +0.01 +0.71 +0.01 +0.00
Qwen3-4B 1003 -0.05 +0.64 -0.05 +0.00
Qwen3-4B 1005 +0.03 +0.62 +0.03 +0.00
Qwen3-4B 1007 -0.01 +0.75 -0.01 +0.00
Qwen3-4B 1009 +0.02 +0.63 +0.02 +0.00
Qwen3-4B 1010 +0.02 +0.69 +0.02 +0.00
Qwen3-4B 1011 -0.05 +0.73 -0.05 +0.00
SmolLM3-3B 1000 -0.04 +0.59 +0.59 +0.00
SmolLM3-3B 1002 +0.01 +0.71 +0.01 +0.00
SmolLM3-3B 1003 -0.05 +0.64 +0.64 +0.00
SmolLM3-3B 1005 +0.03 +0.62 +0.03 +0.00
SmolLM3-3B 1007 -0.01 +0.75 +0.75 +0.00
SmolLM3-3B 1009 +0.02 +0.63 +0.02 +0.00
SmolLM3-3B 1010 +0.02 +0.69 +0.69 +0.00
SmolLM3-3B 1011 -0.05 +0.73 +0.73 +0.00

4.6 Stated reasoning

The models describe what they are doing. Qwen3-4B, on the same instance:

"The industrial_idx shows a strong positive correlation with demand, suggesting that increased industrial activity directly affects electricity demand. The temp_anom and solar_wm2 signals also show strong positive correlations, indicating that temperature and solar irradiance are likely factors influencing demand. The humidity and holiday_flag signals show weaker or negative correlations, which may indicate that these factors are not direct causes but rather proxies or indirect influences."

It excludes two true causes and includes one signal with no causal pathway, and the stated reason is the correlation column.


5. Q3 - what would close the gap

The second experiment localises the problem: it is a data deficit and a structural one at the same time. With interventional evidence supplied, the effect estimates re-anchor on it, but the model still does not update its causal model — it treats the measurement as the answer to one question rather than as a constraint on the whole structure. Three levers follow.

5.1 Data: interventional, not more observational

Causal structure is not identifiable from observational data. The lever is not more data but different data: examples where the do-distribution is known. These are cheap to generate in engineered systems — power grids, HVAC, chemical plants, traffic, industrial batching — where a simulator or a logged randomised trial gives the interventional distribution for free. Low-resource domains currently have essentially none, so a reusable corpus of interventional pairs for agriculture, energy and health would be worth more than another benchmark. Concretely: (observational context, do-effect) triples generated from physically-motivated simulators, released with the generators, so others can generate interventions in their own domain rather than only consuming this one.

5.2 Fine-tuning: the interventional objective, not the causal vocabulary

More causal text makes models better at the phrase causal, not at the quantity. The supervision needed is on the target: given an observational context and a question of the form "what happens to Y if X is set to v", predict the counterfactual outcome. Two choices matter.

  • Counterfactual pairs within one domain. Training on (same context, X set high) and (same context, X set low) side by side forces the model to condition on the intervention rather than on the correlation, because the two examples share a context and differ only in the intervention. This is the main ingredient, and it is free once a simulator exists.
  • A curriculum containing the wrong-correlation case. If every training pair has a positive correlation between cause and effect, the model learns correlation again, with better manners. The training distribution has to include the suppressor, the confounder and the proxy — the three cases where correlation and mechanism disagree.

5.3 Architecture: a separate mechanism channel

A fluent causal claim and a plausible-sounding prior come out of the same token stream, so a wrong mechanism ends up stated as if it had been verified. The fix is to make the claim checkable before it is written:

  • have the model emit a structured intermediate — a candidate causal graph, or a ranked set of candidate mechanisms with predicted signs and magnitudes — in a constrained schema;
  • score it against an interventional criterion (does it reproduce the observed trial measurement? does it survive a refit on held-out data?);
  • let a verifier reject the mechanism before any prose is generated, and fall back to the best-supported answer rather than the most fluent one;
  • let the interventional evidence update the graph, not just one number inside it.

The same discipline appears in a different project: in KernelOPT (arXiv:2609.30059) an LLM agent proposes optimised GPU kernels, and a four-gate verification cascade — static validation, multi-seed correctness, end-to-end float64 fallback verification, and a performance gate — decides whether a kernel may replace the compiler baseline. The verifiable claim is checked and the unverified one is discarded; the same design applies directly to a causal claim.

5.4 Cheapest first experiment

Fine-tune a 1.7B model on synthetic interventional pairs from the simulator in code/dgp.py, then re-run the same three tasks. If the causal-selection F1 moves from 0.480 toward the partial-correlation baseline's 0.857, the gap was data rather than capability, and the remaining work is data engineering rather than new modelling. If it does not move, the problem is architectural and section 5.3 is the lever to pull. Either way the experiment is a few GPU-hours, and the harness is here.


6. Repository layout

code/       dgp.py           structural equation model, trial simulator, instance generation
            run_eval.py      prompts, model loading, generation, parsing
            baselines.py     |corr| / partial-correlation / oracle selectors
            analyze.py       every metric reported above
            make_figures.py  the figures
            verify.py        re-derives every number in this README from scratch
data/       instance_*.csv        the observational data, standardised, one file per instance
            ground_truth_*.csv    per-signal truth: causal effect, correlation, trial measurement
            index.csv             which instance has which randomised trial
results/    raw_*.json            every generation, verbatim, per model
            summary.json          all metrics
            summary_trial.json    all metrics, trial condition
            report.txt            the human-readable report
            case_study.md         the worked example in section 4.5
            verification.txt      the output of verify.py, which re-derives every number here
figures/    fig1..fig3

7. Reproducing

pip install torch transformers accelerate numpy pandas scikit-learn matplotlib
cd code
python run_eval.py --instances 12 --out ../results
python run_eval.py --instances 12 --out ../results \
       --tasks select_trial intervene_trial --merge
python analyze.py
python make_figures.py

A single H200 (71 GB MIG instance) is enough; the study is under an hour of GPU time including downloads. The prompts are in code/run_eval.py and every raw generation is in results/, so each number above can be checked against the model's own words.

8. Limitations

  • The data are synthetic. That is deliberate — exact ground truth requires it — but it bounds the external validity claim, which should be read narrowly.
  • 12 randomised instances per model and 5 models. The effect sizes are large relative to the variance, but this is not a large-sample study.
  • Only 0.6B-6B open-weight models were in scope. Whether frontier-scale models behave differently is open, and is a separate question from whether the benchmark gap exists.
  • The trial condition supplies interventional evidence for half the signals only, so the 'no trial' subset is the available handle on purely observational behaviour.

9. License

MIT. See LICENSE.

Total size
5.15 MB
Files
46
Last updated
Oct 22
Pre-warmed CDN
US EU US EU

Contributors