- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3
Read review
Paper summary
This paper introduces Witness, a benchmark for evaluating dependency analysis in LLM agents. The motivating problem is that information-flow tracking (IFT) defenses against prompt injection require knowing which parts of an agent’s input context each output decision depends on. For conventional programs, such dependencies can often be determined through program analysis, but no analogous reliable mechanism exists for LLM inference. Existing systems either conservatively assume that every LLM output depends on every input, which causes severe label propagation, or rely on approximate dependence estimation. Witness formalizes the task by segmenting an agent trajectory into context units and output units and annotating whether each output unit depends on each context unit. The benchmark distinguishes data dependence from control dependence, with finer-grained subtypes such as supply, compute, filter, etc.
Critical strengths
- The problem is important and clearly motivated. The paper identifies a concrete trust assumption in IFT-based defenses: if the dependency analyzer misses a true dependence on untrusted content, malicious influence may escape tracking. The connection between dependency recall and prompt-injection security is clear and practically relevant.
- The benchmark fills a plausible evaluation gap. The contribution is not another prompt-injection defense, but a benchmark for the dependency-estimation component that several defenses implicitly rely on. I found this framing useful because it isolates a component that can otherwise be hidden inside a larger security system.
- The annotation methodology is careful. The authors define dependence using a relatively concrete counterfactual criterion and distinguish direct value provenance from decision/control dependence. Two annotators independently label the data and achieve high agreement before adjudication.
Critical weaknesses
-
The notion of “ground-truth dependence” is still partly semantic and subjective Although the paper provides a detailed rubric and reports strong inter-annotator agreement, the labels are not ground truth in the same sense as dependencies in a deterministic program. The counterfactual criterion sometimes requires annotators to judge whether an output would “plausibly” change, and the rubric contains explicit exceptions for multi-source decisions and control relations. Though this does not invalidate the benchmark, but I think the paper could more clearly distinguish human semantic dependence from causal dependence inside the model. In particular, an LLM may internally rely on a context unit even when human annotators judge that unit unnecessary under the task-level counterfactual definition.
-
The benchmark is dominated by AgentDojo Although the paper includes 11 agents, AgentDojo contributes 824 of 1,560 output units and 2,880 of 4,715 labeled dependencies, which is a large fraction of the benchmark. This raises a question about how much the aggregate results reflect the diversity of real-world agent frameworks versus one benchmark ecosystem.
Questions for authors
-
How should we interpret the Witness labels relative to causal dependence inside the LLM? Could there be context units that genuinely affect the model’s internal computation but fail the benchmark’s human counterfactual test, or vice versa?
-
Why are reasoning-disabled results used as the primary evaluation? Since Appendix Table 13 shows meaningful recall improvements from reasoning for some models, would the main conclusion change if each model were evaluated under its strongest available inference setting?
2 comments
Response to Reviewer Ahqz (Part 2)
Q2. Why are reasoning-disabled results used as the primary evaluation.
We set temperature to zero in the main evaluation. Some reasoning APIs reject or ignore it, so we disable reasoning to ensure consistent settings across models.
During rebuttal, we evaluated all 12 models with reasoning enabled and their default temperature settings.
Table R3.2 compares reasoning-enabled results with Table 2 of the submission. Most models improve in precision and recall, but key findings hold: dependency recovery remains unreliable, complete sets are recovered for at most half the outputs, and control recall remains much lower than data recall.
Table R3.2. Evaluation results with reasoning enabled (EM: exact match, P: precision, R: recall).
| Model | EM | Overall P | Overall R | Data P | Data R | Control P | Control R |
|---|---|---|---|---|---|---|---|
| Claude-Opus-5 | 0.522(0.076) | 0.810(0.072) | 0.835(-0.006) | 0.899(0.087) | 0.928(-0.001) | 0.771(0.066) | 0.794(-0.008) |
| Claude-Sonnet-5 | 0.518(0.037) | 0.830(0.095) | 0.739(-0.010) | 0.907(0.050) | 0.861(-0.002) | 0.792(0.111) | 0.684(-0.014) |
| Claude-Haiku-4.5 | 0.506(0.138) | 0.867(0.087) | 0.698(0.144) | 0.904(0.134) | 0.863(0.113) | 0.845(0.058) | 0.625(0.158) |
| GPT-5.6-Sol | 0.445(-0.071) | 0.697(-0.122) | 0.867(0.124) | 0.877(0.034) | 0.936(0.059) | 0.632(-0.174) | 0.836(0.153) |
| GPT-5.6-Terra | 0.475(-0.004) | 0.717(-0.117) | 0.838(0.171) | 0.871(0.018) | 0.920(0.084) | 0.658(-0.165) | 0.801(0.209) |
| GPT-5.6-Luna | 0.425(0.019) | 0.733(0.016) | 0.817(0.165) | 0.888(0.147) | 0.915(0.097) | 0.672(-0.032) | 0.772(0.195) |
| DeepSeek-V4.1-Flash | 0.520(0.075) | 0.831(0.056) | 0.779(0.092) | 0.865(0.125) | 0.904(0.040) | 0.812(0.014) | 0.722(0.115) |
| DeepSeek-V4-Pro | 0.492(0.081) | 0.754(0.153) | 0.760(0.020) | 0.841(0.218) | 0.895(0.015) | 0.712(0.123) | 0.699(0.022) |
| DeepSeek-V4-Flash | 0.470(0.090) | 0.767(0.132) | 0.790(0.106) | 0.863(0.333) | 0.916(0.058) | 0.722(-0.005) | 0.733(0.127) |
| Qwen3.8-27B | 0.417(-0.048) | 0.694(-0.116) | 0.703(0.040) | 0.782(-0.045) | 0.847(0.036) | 0.650(-0.150) | 0.638(0.042) |
| Qwen3.5-27B | 0.491(0.053) | 0.787(0.104) | 0.691(0.022) | 0.801(0.094) | 0.844(0.024) | 0.778(0.109) | 0.622(0.021) |
| Qwen3.5-9B | 0.307(0.085) | 0.612(0.187) | 0.557(0.062) | 0.685(0.268) | 0.634(-0.057) | 0.578(0.146) | 0.522(0.115) |
Cells show value (Δ). Δ = reasoning enabled − reasoning disabled.
Further Clarification of the Identified Weakness
W1. The paper should more clearly distinguish human semantic dependence from causal dependence inside the model.
See Q1.
W2. AgentDojo contributes a large fraction of the benchmark. The aggregate results may not reflect the diversity of real-world agent frameworks.
We acknowledge that AgentDojo contributes over half of the dependency relations. However, excluding AgentDojo leaves many patterns and observations unchanged.
As shown in Table 1 of the submission, subtype distributions remain similar after excluding AgentDojo, except for Filter and Rank.
Table R3.3 reports scores after excluding AgentDojo and differences from Table 2 of the submission. Most changes are small, and many findings hold: dependency identification remains unreliable, models have lower control recall than data recall, and more capable models perform better.
The remaining 10 agents cover common tasks: question answering, coding, and personal assistance. These results suggest our findings reflect broader patterns in real-world agent frameworks rather than being specific to AgentDojo.
Table R3.3. Dependence attribution excluding AgentDojo (EM: exact match, P: precision, R: recall).
| Model | EM | Overall P | Overall R | Data P | Data R | Control P | Control R |
|---|---|---|---|---|---|---|---|
| Claude-Opus-5 | 0.410(-0.034) | 0.726(-0.012) | 0.834(-0.002) | 0.783(-0.016) | 0.870(-0.058) | 0.692(-0.018) | 0.811(0.016) |
| Claude-Sonnet-5 | 0.492(0.010) | 0.786(0.039) | 0.735(-0.011) | 0.817(-0.051) | 0.796(-0.064) | 0.766(0.072) | 0.697(0.002) |
| Claude-Haiku-4.5 | 0.374(0.001) | 0.789(0.011) | 0.581(0.029) | 0.777(0.011) | 0.670(-0.073) | 0.798(0.012) | 0.525(0.059) |
| GPT-5.6-Sol | 0.519(0.005) | 0.824(0.001) | 0.759(0.016) | 0.872(0.029) | 0.808(-0.069) | 0.793(-0.018) | 0.729(0.046) |
| GPT-5.6-Terra | 0.479(0.004) | 0.821(-0.018) | 0.705(0.039) | 0.855(-0.005) | 0.747(-0.075) | 0.798(-0.027) | 0.679(0.083) |
| GPT-5.6-Luna | 0.394(-0.009) | 0.667(-0.046) | 0.691(0.045) | 0.701(-0.048) | 0.766(-0.043) | 0.644(-0.048) | 0.644(0.071) |
| Qwen3.8-27B | 0.481(0.018) | 0.851(0.046) | 0.659(-0.005) | 0.848(0.035) | 0.717(-0.098) | 0.852(0.052) | 0.623(0.026) |
| Qwen3.5-27B | 0.450(0.010) | 0.760(0.075) | 0.677(0.010) | 0.770(0.060) | 0.756(-0.065) | 0.753(0.082) | 0.628(0.029) |
| DeepSeek-V4-Pro | 0.446(0.034) | 0.651(0.050) | 0.770(0.031) | 0.646(0.022) | 0.854(-0.026) | 0.655(0.066) | 0.718(0.042) |
| DeepSeek-V4.1-Flash | 0.463(0.019) | 0.747(-0.021) | 0.724(0.037) | 0.746(0.020) | 0.797(-0.064) | 0.748(-0.050) | 0.679(0.070) |
| DeepSeek-V4-Flash | 0.387(0.003) | 0.651(-0.009) | 0.713(0.034) | 0.608(0.057) | 0.815(-0.029) | 0.690(-0.065) | 0.650(0.045) |
| Qwen3.5-9B | 0.210(-0.013) | 0.361(-0.063) | 0.465(-0.030) | 0.359(-0.059) | 0.603(-0.089) | 0.366(-0.065) | 0.380(-0.027) |
Cells show value (Δ). Δ = excluding AgentDojo − full benchmark.
Response to Reviewer Ahqz (Part 1)
We thank the reviewer for the constructive feedback. We address the concerns and questions below.
Answers to Required Questions:
Q1. How should we interpret the Witness labels relative to causal dependence inside the LLM?
Witness labels are based on human judgments about task semantics. Many of these labeled dependencies also affect the model’s internal computation. We support this claim with two experiments: direct input interventions and context masking.
For direct input interventions, we sample labeled dependency pairs (c, o), where o is performed at step t. We modify c according to the dependency subtype, then provide the modified trajectory before step t to an LLM, and check whether o changes in its output. We use the following modification rules:
We sample more than 15% of the labeled relations in each subtype. We use GPT-5.6-Luna for both modifications and validation. We request GPT-5.6-Luna to analyze each trajectory three times and consider a dependency is validated only if o changes in all three runs.
Overall, the input interventions confirm 645 of the 732 sampled relations (88.11%). The breakdown by subtype is shown in Table R3.1.
We conduct a similar experiment to validate independent annotations. We sample pairs (c, o) where c appears in the trajectory before o is performed and has no labeled relation to o. We use GPT-5.6-Luna to modify c, then ask the same LLM to analyze the modified trajectory three times. We confirm an independent annotation only if o remains unchanged in all three runs. Overall, 667 of the 732 sampled (91.12%) pairs met this criterion.
Table R3.1. Validating the consistency between human-annotated dependencies and actual model behavior.
For context masking, we follow Sia et al. [1]. For each output unit o, we provide the preceding trajectory to the model and mask attention to all tokens corresponding to o’s annotated dependent units. We then check whether o changes in the model’s output.
We use Qwen3-8B because its open implementation and standard self-attention architecture make context masking straightforward to apply across all layers and attention heads.
We first identify steps that Qwen3-8B can reproduce without masking. For each of the 704 steps in Witness-Gold, we provide the preceding trajectory to the model three times. We retain the 166 steps that are reproduced in all three runs. These steps contain 342 output units: 120 with only data dependencies, 164 with only control dependencies, and 58 with both.
For each retained output unit, we jointly mask all its annotated context units and repeat the experiment three times. We count an output unit as affected only if it changes in all three runs.
Overall, masking changes 117/120 output units with only data dependencies, 154/164 with only control dependencies, and 50/58 with both. These results suggest that the annotated context units influence the model’s internal computation when generating these outputs.
[1] Where does In-context Learning Happen in Large Language Models?, NeurIPS’2024.