Skip to content
EventPast event

The paper proposes using a language model to simulate tool responses for control evaluation.

4 reports4 reporting sourcesLatest report

Start here

AI overview

A paper published on October 23, 2026 proposes a "simulated control evaluation" method: using a language model to simulate tool responses and checking predictions in the actually executable LinuxArena environment, the root mean square error of safety predictions is 3.6 percentage points. This result is based on five monitors, a 5% audit budget and an independently sampled task set; the corresponding error for the control environment BashArena is 10.1 percentage points. The paper notes that this is not a general error bound for all deployment environments, and that policies discovered in simulation also improved monitoring performance on held-out LinuxArena attacks. On October 24, Redwood Research published a paper introducing simulated control evaluation (SCE): using an LLM environment simulator instead of real execution, letting attack policies continue the task on agent transcripts from real codebases, with tool calls not actually executed, tool outputs generated by the simulator, and an attack success judgment model then used to estimate whether the attack succeeded; on LinuxArena the prediction error for the main safety metric is 3.6 percentage points, and on BashArena 10.1 percentage points. Redwood Research's Buck Shlegeris reposted the paper and commented that the approach works well and makes it much more feasible to test control measures in the deployment scenarios that actually matter.

Generated by AI from reports · Updated 22 days ago

Attention and developments

Attention, hourly

Now 42, 0% in 24 hours·Peak 42 (Nov 13 03:00)

Nov 14Nov 15Nov 16Nov 17Nov 18Nov 19Nov 20Now 42

Progress timeline

Oct 25
Oct 24
  1. X @bshlgrs
    Research shows that using LLM simulators in place of real environments to evaluate the effectiveness of AI control protocols works well.

    A new paper proposes that large language model simulators can replace real environments to evaluate AI control protocols, testing whether monitors can prevent agents from causing harm during deployment; the paper states that the models can simulate control evaluation environments for specific deployment scenarios with relatively high accuracy. Buck Shlegeris of Redwood Research shared the paper and commented that the approach works well, making it much more feasible to test control measures in the deployment scenarios that are actually of concern.

  2. LessWrongSelected
    Simulated control evaluations (SCEs): using an LLM simulator to bring control evaluations closer to real deployment

    The authors release a paper on simulated control evaluations (SCEs), proposing to use an LLM simulator to fabricate tool returns, testing without actually executing an attack whether monitors that developers use to supervise agents can detect models that deliberately insert vulnerabilities, steal credentials or sabotage experiments. The method starts from an agent's work trajectories in a real codebase, with gpt-6-astra simulating tool outputs (with gpt-5.6-sol as fallback), and gpt-5.6-sol simultaneously serving as the attack policy and the judge of attack success. The paper reports that, validated against real execution on LinuxArena, the gap between simulated and real is comparable to the gap between two independent real runs; under a 5% audit budget, the safety prediction RMSE is 3.6 percentage points, and for BashArena it is 10.1 percentage points; SCE explains 95% of the expected safety differences between side-task categories and 78% between environments, and sampling tasks at 4.6% of deployment scale already exceeds BashArena's full prediction.

  3. Redwood ResearchSelected
    Redwood Research publishes paper on the Simulated Control Evaluation (SCE) method

    Redwood Research published a paper proposing Simulated Control Evaluation (SCE), which uses an LLM environment simulator instead of real execution to test whether monitoring deployed by developers can intercept an agent before it carries out harmful behavior. The method is to have an attack policy continue the task on agent transcripts from real codebases, with tool calls not actually executed, tool outputs generated by the simulator, and then an attack-success judgment model used to estimate whether the attack succeeded. On LinuxArena, where real execution is possible, SCE's prediction error for the main safety metric was 3.6 percentage points, and on BashArena it was 10.1; the difference in behavior distribution between simulated trajectories and real execution was close to the difference between two independent real executions. The authors say the simulated sample size is less than 5% of deployment, that sampling only 4.6% of tasks is enough to exceed the safety prediction of the full BashArena, and note that fidelity can only be verified in environments where real execution is possible, and that there is not yet a simulator for long-horizon multi-agent attacks.

Oct 23
  1. Paper Tracker
    Simulated Control Evals

    The paper proposes simulated control evaluations (SCEs), using language models to simulate tool responses in place of real environments, thereby conducting control evaluations in more environments. The authors tested on LinuxArena, which can provide real execution results; under a 5% audit budget and five monitors, SCE's root mean square error in predicting safety was 3.6 percentage points, and for BashArena it was 10.1. The KL divergence between SCE and real execution is more than ten times lower than that of BashArena, and it can explain 78% to 85% of the differences in attack success rate, monitor score, and safety across environments, requiring fewer than 5% of the deployment task set's task pairs as samples. Policies discovered in simulation also improved monitoring performance on held-out LinuxArena attacks, with an effect close to the improvement brought by real-execution policies.

Related discussion

Total 17 items

Attention trend

Hourly attention

Current attention 42·Peak in the comparable period 42 (Nov 13 03:00)·Change over the comparable past 24 hours 0%

0204060Nov 1303:00Nov 1511:00Nov 1718:00Nov 2002:00