acceptodds
Under review as a conference paper at ICLR 2027

GIF: Locally Sound Geometric Information Flow Control for LLMs

Abstract

Large language models increasingly mediate interactions between sensitive data, untrusted inputs, and privileged actions, creating security and privacy risks in agentic systems. Information Flow Control (IFC) offers a promising defense, but existing approaches lack a principled way to reason about information flow through the model itself and therefore suffer from severe taint explosion. We introduce Geometric Information Flow (GIF), a semantic framework that uses the LLM's Jacobian and local output geometry to measure how input spans influence model outputs. GIF upper-bounds the mutual information induced by local input perturbations and admits scalable estimation via automatic differentiation and low-rank approximation. Across integrity and confidentiality benchmarks spanning prompt injection and privacy leakage, GIF consistently outperforms token-level attribution baselines and, when paired with lightweight declassifiers, matches or exceeds direct LLM-as-judge baselines while using up to lower model inference cost. GIF also transfers across model scales and families, including from surrogate models up to smaller than their targets. Our results establish geometric information flow as a principled and practical foundation for scalable IFC in LLM systems.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
information flow control, LLM, security, privacy

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_LuMoOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GIF models local information flow from an input span to the next-token distribution using the model Jacobian and softmax Fisher geometry. It derives a local Gaussian-channel information bound, uses the Fisher-pullback trace as an influence proxy, and estimates that trace with randomized matrix-vector operations. Experiments test source localization and the use of selected spans for policy decisions on three agent-security benchmarks, including approximations and surrogate analysis models.

Critical strengths

The local information-theoretic result gives a precise interpretation to the influence score. The fixed-context, small-perturbation assumptions are explicit, and Appendix B develops the argument from expected KL to the Gaussian-channel proxy. Connecting this result to an estimator that avoids materializing the full Jacobian is a useful contribution for studying information flow in language models.

The empirical evaluation connects the estimator to practical source selection. Table 2 compares sketched and exact influence, and the same-model comparisons in Figure 2 show that compact attributed evidence can preserve or improve policy decisions while reducing the judge's input.

Critical weaknesses
  1. The headline claim of matching GPT-5.5 xhigh at more than 81 times lower cost is not supported by a comparison on a common evaluation population. Section 4.2 and Figure 2 compare open-model rows evaluated on each model's own trajectories with a GPT-5.5 row pooled across all six agent models, as the caption states. Differences in trace difficulty and violation prevalence therefore enter the quality comparison. The shared-sample attribution experiment in Appendix F.2 concerns a different task and does not repair this policy-decision comparison. Report both systems on the same trajectories, actions, and labels, and include GIF's gradient and ranking costs in the total. The paper should distinguish the cost of the policy judge from the complete attribution-and-decision pipeline.

  2. The incremental contribution of the Fisher-based selector needs a more precise account. JacobianScopes/FisherScope already defines the softmax Fisher pullback and its trace as token influence (January 2026 version, Section 3.2 and Appendix A). The additional contribution here is the local information bound, estimation procedure, and policy application. For the application, Figure 2 does not compare GIF-selected evidence with BM25 or GIF-minus evidence supplied to the same judge at the same length and format. This matters because Table 1 shows that BM25 exceeds GIF on average MSB localization at @15: AUROC 0.92 versus 0.89 and AUPRC 0.94 versus 0.91. A matched downstream comparison would establish whether the geometric selection itself improves policy decisions beyond compact evidence selection.

Questions for authors
  1. What are the policy-decision results and full costs when the Qwen-GIF and GPT-5.5 systems receive exactly the same trajectory/action examples? Please distinguish attribution computation from judge inference in the reported cost factor.

  2. With the judge, evidence length, and prompt format fixed, how do GIF, GIF-minus, and BM25 span selection compare? Please also state precisely which mathematical result is new relative to the Fisher-pullback construction in JacobianScopes/FisherScope.

3 comments

We thank the reviewer for their thoughtful assessment and recognition of GIF's theoretical and practical contributions. We address both concerns below.

Q1: GPT-5.5 comparison and computational cost

We clarify that the 81× figure explicitly refers to downstream model inference cost, not total pipeline cost, as stated in Section 4.2. It illustrates that GIF enables a substantially smaller declassifier (Qwen3-8B) to achieve detection performance comparable to GPT-5.5 xhigh, rather than claiming an 81× end-to-end cost reduction.

We acknowledge that the cross-model comparison uses different trajectory populations and is therefore illustrative rather than a strictly paired comparison. Importantly, Figure 2 also provides same-model, same-trajectory comparisons, demonstrating that GIF-guided declassification preserves or improves policy-decision quality while reducing downstream token consumption by 59–73% on AgentDojo and MSB.

GIF's computational overhead is modest, requiring a single backward pass per analyzed token through the model being analyzed. Only a small number of output tokens need to be analyzed, corresponding to security-relevant actions such as generated tool-call names and parameters. Moreover, Figure 3 demonstrates that a lightweight 0.6B surrogate preserves downstream detection performance even for trajectory-generating models up to 200× larger, enabling efficient information-flow analysis even for substantially larger trajectory-generating models.

Q2: Incremental contribution and downstream evidence-selection baselines

We acknowledge that Liu et al. (2026) also employ the Fisher pullback and its trace for token-level attribution. However, GIF's primary contribution is not the Fisher pullback itself, but the information-theoretic framework and formal guarantees constructed around it.

Specifically, we introduce a local Gaussian surrogate channel whose KL geometry (Section 2.2) matches that of the autoregressive model, yielding an explicit mutual-information measure: CapT(τ;x)=12log⁡det⁡(I+τMT(x)).\mathrm{Cap}_T(\tau;x)=\frac{1}{2}\log\det(I+\tau M_T(x)).

We establish direction-uniform second-order faithfulness (Theorem 2.2) and prove that this measure upper-bounds true perturbation-to-output mutual information to first order (Theorem 2.3): I(UT;Y)≤CapT(τ;x)+o(τ).I(U_T;Y)\leq\mathrm{Cap}_T(\tau;x)+o(\tau).

We further develop scalable matrix-free estimation using Hutch++ and connect its estimation guarantees to our soundness theorem. Neither the information-channel construction nor these soundness results are provided by Liu et al., whose objective is token-level attribution rather than quantitative information-flow semantics. We already cite Liu et al. as closely related work and will further clarify these distinctions in the revised paper.

Regarding downstream baselines, our evaluation deliberately separates attribution quality (Section 4.1) from policy-decision quality (Section 4.2). Section 4.1 directly compares GIF against GIF−, RTBAS-LLA, BM25, and random selection without an LLM judge, isolating the attribution signal. Section 4.2 establishes that GIF-selected evidence supports accurate semantic policy decisions with substantially reduced context.

Evaluating every alternative selector with the same downstream judge would constitute an additional end-to-end systems comparison. Our claims concern GIF's information-flow measurement and its effectiveness in supporting downstream policy decisions, not superiority over every possible evidence-selection pipeline.

Thank you for clarifying the cost figure and the relation to Liu et al. I am keeping my rating at 3. The local information-theoretic result and matrix-free estimation remain the main reasons for my recommendation.

The acknowledgment that the 81x comparison uses different trajectory populations is important. Restricting the number to downstream inference cost does not resolve the quality-comparison issue. Please describe this comparison as illustrative throughout the paper, including the abstract, and distinguish it from the same-model results. A backward-pass count and surrogate-transfer results also do not quantify the attribution overhead of the evaluated pipeline. A matched comparison needs the same trajectories, actions, and labels, with attribution and judging costs reported separately and in total.

The downstream selector question remains unanswered. Section 4.1 measures localization; it does not establish what GIF's ranking adds to policy decisions. A focused comparison of GIF, GIF-minus, and BM25 with the same judge, trajectories, evidence length, and prompt format would distinguish the contribution of the geometric ranking from the benefit of shortening the context. This is a targeted control for the application claim.

We thank the reviewer for their follow-up and continued recognition of GIF's theoretical and methodological contributions.

Regarding the GPT-5.5 comparison, we appreciate the reviewer's suggestion to control for differences in trajectory populations. We have disaggregated our results to compare GIF-guided Qwen3-8B against GPT-5.5 xhigh on identical trajectories, actions, and ground-truth labels across all six trajectory-generating models on AgentDojo and MSB, the two benchmarks underlying our 81× comparison.

The table below reports F1 (%) for GIF-3 with Qwen3-8B as the downstream declassifier versus GPT-5.5 xhigh performing full-trajectory judging.

ModelAgentDojoAgentDojoMSBMSB
GIF-3 + Qwen3-8BGPT-5.5 xhighGIF-3 + Qwen3-8BGPT-5.5 xhigh
Qwen3-8B74.076.183.184.0
Qwen3-14B89.790.591.592.4
Qwen3-32B88.188.482.284.2
Gemma-4-31B86.690.161.283.6
GPT-OSS-20B72.775.576.877.8
GPT-OSS-120B77.374.987.489.2
Macro-average81.482.680.485.2

These matched results demonstrate that GIF-guided Qwen3-8B achieves comparable detection performance on AgentDojo (81.4% vs. 82.6% F1) and MSB (80.4% vs. 85.2% F1). On MSB, the difference is predominantly driven by the Gemma-4-31B configuration; across the other five models, performance differs by approximately 1–2 percentage points. We will include this disaggregated comparison in the revised manuscript and clarify that the 81× ratio refers specifically to downstream declassifier inference cost, rather than total pipeline cost.

Regarding computational overhead, we agree that detailed end-to-end cost analysis would be valuable for practical deployment. Our current work characterizes GIF's computational requirements, while further implementation optimization and deployment-level benchmarking are natural directions for future work.

Regarding downstream evidence-selection controls, we thank the reviewer for this constructive suggestion. A systematic comparison of alternative evidence selectors under a fixed declassifier would provide additional insight into their relative effectiveness in downstream policy decisions. While our current evaluation focuses on information-flow measurement and demonstrating its applicability to semantic declassification, we agree that exploring alternative evidence-selection and policy-decision mechanisms is a valuable direction for future work.

Reviewer_scapOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Geometric Information Flow (GIF), a quantitative measure of how much an input span influences an LLM's output. It is intended as a basis for information flow control (IFC) in agentic systems, where untrusted or confidential context can affect actions. GIF pulls the softmax Fisher metric of the next-token distribution back through the model's span-to-hidden-state Jacobian. It then defines flow as the mutual information of a Gaussian channel that matches this local KL geometry, ½ log det(I + τM). Under a local regularity assumption, the authors prove, and formalize in Lean 4, that GIF upper-bounds the true perturbation-to-output mutual information to first order. In practice they use the trace of the Fisher pullback, estimated with Hutch++ through Jacobian-vector and vector-Jacobian products. Experiments cover AgentDojo and MSB (prompt injection) and AgentDAM (privacy leakage) with six open-weight models. GIF generally ranks annotated policy-relevant tokens better than an attention baseline and a logit-gradient ablation. When GIF selects the top-k evidence spans for an LLM declassifier, detection roughly matches or exceeds full-trajectory judging while cutting tokens by up to 73%. On AgentDojo, GIF scores computed with a 0.6B surrogate preserve downstream performance for targets up to 120B, and GIF remains informative under encoded or paraphrased injections.

Critical strengths
  1. A principled, formally verified notion of flow through the model. The paper ties the Jacobian and output Fisher geometry to an explicit information channel, with a local soundness theorem formalized in Lean 4. This is unusually rigorous for LLM security. It gives IFC a quantitative definition of flow through the model itself, rather than heuristic attribution.

  2. Scalable, practical estimation. Matrix-free Hutch++ with JVP/VJP products avoids materializing the Jacobian. Low-rank sketches already preserve source rankings at rank 1 (Spearman 0.957). A 0.6B surrogate changes downstream F1 by at most 0.7 pp for targets up to 120B, which makes GIF usable when target gradients are unavailable or too costly.

  3. Useful for downstream policy decisions. As an evidence selector for LLM declassifiers, GIF roughly matches or exceeds full-trajectory judges while using 59–73% fewer tokens on AgentDojo and MSB. It notably strengthens small judges: Qwen3-8B gains 9–12 F1 points and matches GPT-5.5 on AgentDojo and MSB. GIF also outperforms BM25 under obfuscated injections.

Critical weaknesses
  1. The downstream evaluation lacks evidence-selection controls. Fig. 2 compares GIF-selected evidence only against full-trajectory judging. The declassifier is never given the top-k spans chosen by BM25, attention (RTBAS-LLA), GIF−, or random selection. It is therefore unclear whether the gains and token savings come from GIF's ranking or simply from shortening the context. The surrogate result reinforces this concern: replacing the 120B target with a 0.6B model changes downstream F1 by at most 0.7 pp. This suggests either that the downstream metric is insensitive to attribution quality, or that GIF captures text-level salience rather than flow through the target model, contrary to the paper's framing.

  2. The gains are marginal and the claims overstated. The abstract states that GIF "consistently outperforms token-level attribution baselines." However, its gain over GIF−, which uses the same gradients on the realized-token logit, is only about 2 AUROC points on average (AgentDojo: 0.75/0.82/0.85 vs. 0.73/0.79/0.83). GIF− is better on MSB for Qwen-C at every cutoff and on the MSB average at @15. The lexical BM25 baseline also outperforms GIF on the MSB average (AUROC 0.92 vs. 0.89 and AUPRC 0.94 vs. 0.91 at @15), and on Gemma4 for both AgentDojo and MSB. The Fisher-geometric machinery therefore adds little practical value over a plain gradient signal.

  3. The confidentiality results are weak. On AgentDAM, average AUPRC is only 0.37–0.40, against 0.26–0.27 for random spans. For Gemma4, GIF's AUPRC (0.12–0.14) falls below random (0.16–0.17). For GPT-OSS-20B, the attention baseline outperforms GIF (AUPRC 0.43–0.51 vs. 0.37–0.45). Confidentiality is one of the paper's two motivating applications, so the supporting evidence is insufficient.

  4. There are no adaptive attacks. All evaluated injections come from the benchmarks' fixed attack sets. The obfuscation study (base64, hexadecimal, spacing, paraphrasing) is not designed against GIF either. An attacker who knows about GIF could, for example:

    • craft injections that drive the target action with near-deterministic outputs, where the softmax Fisher metric, and hence GIF, collapses;
    • route the payload through the model's own reasoning tokens, which GIF treats as fixed context.

    Adaptive attacks are known to break prompt-injection defenses [1, 2], so the robustness of GIF-based enforcement remains unclear.

References [1] Zhan et al., "Adaptive Attacks Break Defenses Against Indirect Prompt Injection Attacks on LLM Agents," arXiv:2503.00061, 2025. [2] Carlini et al., "On Evaluating Adversarial Robustness," arXiv:1902.06705, 2019.

Questions for authors

see weakness

2 comments

We thank the reviewer for their thoughtful assessment and recognition of GIF's theoretical foundations, scalable estimation, and practical applicability. We address the four concerns below.

Q1: Downstream evidence-selection controls and surrogate attribution

Our evaluation deliberately separates attribution quality (Section 4.1) from downstream policy-decision quality (Section 4.2).

Section 4.1 directly compares GIF against GIF−, RTBAS-LLA, BM25, and random selection using PSP@k, AUROC, and AUPRC, independently of any LLM judge. These experiments isolate the quality of GIF's attribution signal. Section 4.2 addresses a complementary question: whether GIF-selected evidence supports accurate semantic policy decisions compared with full-trajectory judging. Running every alternative attribution method through the declassifier would constitute an additional end-to-end system comparison, rather than being necessary to establish either result.

The surrogate experiment demonstrates GIF's practicality in the end-to-end policy-decision pipeline: using a much smaller analysis model preserves downstream detection performance while substantially reducing attribution cost. Separately, our obfuscation experiments show that GIF outperforms lexical BM25, demonstrating that its signal goes beyond textual similarity.

Q2: Magnitude and consistency of attribution improvements

We disagree that the improvements are marginal when considered across the full evaluation. At PSP@15, GIF improves average AUROC over RTBAS-LLA by 16 points on AgentDojo, 28 points on MSB, and 14 points on AgentDAM. GIF also improves over GIF− on AgentDojo (0.85 vs. 0.83 AUROC) and AgentDAM (0.40 vs. 0.36 AUPRC).

We acknowledge that individual baselines outperform GIF in certain configurations, particularly BM25 on MSB, where attacks frequently involve verbatim copying of paths and tool arguments. However, GIF substantially outperforms BM25 under prompt obfuscation (AUROC/AUPRC: 0.649/0.471 vs. 0.473/0.371), demonstrating its value beyond lexical matching.

Our claims concern overall performance across diverse benchmarks and models, not strict dominance in every configuration. We will clarify the wording accordingly.

Q3: Confidentiality results

We disagree that the confidentiality results are weak. GIF achieves the strongest overall performance on AgentDAM across six models, with average AUROC of 0.64–0.69 and AUPRC reaching 0.40, compared with 0.32 for RTBAS-LLA and 0.27 for random selection at PSP@15.

The reviewer highlights a small number of exceptions, primarily GPT-OSS-20B and Gemma 4, which are not representative of the overall results. GIF also improves over GIF− in confidentiality AUPRC, supporting the contribution of distribution-aware geometry.

Furthermore, Section 4.2 demonstrates that GIF-selected evidence supports competitive downstream confidentiality-violation detection on AgentDAM.

Q4: Adaptive attacks and scope of the security guarantee

We acknowledge that our evaluation does not include attacks specifically optimized to evade GIF. However, GIF is not merely a heuristic attribution method: its information-flow bounds are formally verified and provide a principled foundation for security enforcement.

In particular, our local soundness theorem establishes that true perturbation-to-output mutual information is upper-bounded by GIF to first order. Combined with our probabilistic estimation guarantee, this means that sufficiently large local information flow cannot be underestimated by the conservative bound, within the stated assumptions and approximation guarantees. This is an important distinction from heuristic attribution methods, which provide no corresponding soundness guarantee.

The reviewer raises two plausible potential failure modes. Near-deterministic outputs can reduce local Fisher sensitivity, but this does not invalidate our local soundness guarantee. Similarly, our current implementation does not propagate provenance through intermediate reasoning tokens, as acknowledged in Appendix D.2, although the framework permits recursive composition of attribution across intermediate states.

We agree that adaptive robustness warrants further empirical investigation, but the proposed scenarios do not negate GIF's formal guarantees or establish a fundamental weakness of the framework.

We are happy to answer any remaining questions in the discussion period. If our rebuttal has addressed your concerns, we would appreciate a score update.

Reviewer_W177OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Geometric Information Flow (GIF), a framework that approximates the mutual information between input perturbations and the output of an LLM. The authors demonstrate how to efficiently estimate GIF and evaluate its performance on integrity and confidentiality tasks through experimentation. They claim that existing Information Flow Control techniques for LLMs are ad hoc and use model-based heuristics. In contrast, GIF provably describes information flow for LLMs under standard local smoothness assumptions.

Critical strengths

Efficiency. GIF identifies potentially high-risk input tokens, which are then passed to an LLM-based judge for further investigation. Experiments suggest that this procedure is more efficient than full-trajectory judging.

Critical weaknesses

Limited methodological novelty. The idea behind GIF is to estimate the KL di- vergence between the next-token distributions with respect to the perturbed and unperturbed inputs using a Taylor approximation and a Hutchinson estimator of the trace of the Fisher information matrix. The technical novelty is unclear. Specifically, Liu et al. (2026) already use Fisher pullback trace for single-token attribution.

Insufficient comparison. The performance comparison does not seem convincing. Only one alternative method Zhong et al. (2025) for estimating input-output influence is listed in Table 1. The other two are a lexical-relevance baseline and a random-ranking baseline, respectively.

Insufficient Motivation. The connection to security-relevant dependence remains insufficiently estab- lished. The paper does not clarify how much GIF’s scores contribute to detecting policy violations beyond using the LLM judge.

Unnecessary Gaussian-channel construction. It is unclear why the Gaussian channel is introduced. The KL divergence-to-trace link can be studied without this intermediate step.

==================================

Toni J. B. Liu, Baran Zadeo˘glu, Nicolas Boull´e, Rapha¨el Sarfati, and Christopher J. Earls. Jacobian scopes: Token-level causal attributions in LLMs, 2026. URL https://arxiv.org/abs/2601.16407v1.

Peter Yong Zhong, Siyuan Chen, Ruiqi Wang, McKenna McCall, Ben L. Titzer, Heather Miller, and Phillip B. Gibbons. RTBAS: Defending LLM agent

Questions for authors

Could you identify the technical novelty of GIF, especially compared to Liu et al. (2026)?

Why were additional gradient-based and intervention-based attribution methods like Li et al. (2016); Sundararajan et al. (2017) not included in the comparison?

Can GIF be used to enforce confidentiality and integrity policies without an LLM-based judge? What does the Gaussian-channel construction contribute?

In particular, which theoretical guarantee or interpretation would be lost if this construction were omitted?

==================================

Toni J. B. Liu, Baran Zadeo˘glu, Nicolas Boull´e, Rapha¨el Sarfati, and Christopher J. Earls. Jacobian scopes: Token-level causal attributions in LLMs, 2026. URL https://arxiv.org/abs/2601.16407v1.

Jiwei Li, Will Monroe, and Dan Jurafsky. Understanding neural networks through representation erasure, 2016. URL https://arxiv.org/abs/1612.08220.

Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Re- search, pages 3319–3328. PMLR, 2017. URL https://proceedings.mlr.press/v70/sundararajan17a. html.

3 comments

We thank the reviewer for their feedback and the opportunity to clarify GIF's theoretical contributions and experimental methodology. We address each question below.

Q1: Could you identify the technical novelty of GIF, especially compared to Liu et al. (2026)?

GIF's central contribution is a formally verified, information-theoretic framework for quantifying information flow through LLMs, with provable soundness guarantees and scalable estimation. Unlike token attribution, which identifies influential inputs, GIF quantifies information transmitted from an input span to an LLM's output and establishes a provable relationship between true information flow and a computable measure.

While Liu et al. (2026) also use Fisher geometry and establish its relationship to local KL divergence, their measure characterizes local distributional sensitivity, not mutual information. This distinction is particularly important for our target security-sensitive applications. Unlike best-effort attribution, where approximate identification of influential tokens may be sufficient, information-flow security requires sound guarantees: underestimating information leakage can cause actual security violations to go undetected. GIF addresses this fundamental limitation by establishing a provable upper bound on true perturbation-to-output mutual information, rather than relying solely on empirical attribution scores.

Specifically, Liu et al. define their "total information" through expected KL divergence under isotropic input perturbations (Appendix A.11, Eq. 21, p. 25) and describe Fisher Scope as a rank-one approximation of this quantity (Eq. 23, p. 26). They do not establish any bound on mutual information between input perturbations and model outputs. Their theoretical results are valid for their attribution objective, but do not provide an information-flow soundness guarantee.

Our technical contributions are threefold:

  1. Information-theoretic surrogate channel (Section 2.2). We construct a local Gaussian surrogate channel whose KL geometry matches that of the LLM, yielding a closed-form information-theoretic measure CapT(τ;x)=12log⁡det⁡(I+τMT(x)).\mathrm{Cap}_T(\tau;x)=\frac12\log\det(I+\tau M_T(x)). Crucially, we prove that this quantity is exactly the mutual information of the constructed surrogate channel (Eq. 4, Theorem B.5), rather than a heuristic sensitivity or attribution score.

  2. Formally verified soundness guarantees (Theorems 2.2 and 2.3). We establish direction-uniform second-order faithfulness and prove I(UT;Y)≤CapT(τ;x)+o(τ).I(U_T;Y)\leq\mathrm{Cap}_T(\tau;x)+o(\tau). Under our local regularity assumptions, GIF therefore does not underestimate true perturbation-to-output mutual information to first order. Establishing this result requires combining a variational mutual-information bound, Gaussian perturbation averaging, and the exact mutual-information characterization of the surrogate channel. Our theoretical development is further supported by Lean mechanization. Liu et al. provide neither this surrogate-channel construction nor a corresponding information-flow soundness guarantee.

  3. Scalable estimation with probabilistic soundness (Section 3). We develop a matrix-free Hutch++ estimator that avoids explicit Jacobian construction and connect its estimation-error bounds to our local soundness theorem, obtaining a probabilistic soundness guarantee (Corollary C.2). We demonstrate its applicability to both integrity and confidentiality analysis across multiple agent benchmarks and model families.

In summary, Liu et al. use Fisher geometry for best-effort token attribution; GIF uses it to construct an information-theoretic framework for measuring information flow with formally established soundness guarantees. The novelty is not the Fisher geometry itself, but the surrogate-channel construction, its provable relationship to true mutual information, and its scalable estimation with probabilistic guarantees. These guarantees are particularly consequential for security applications, where missed information flows can directly translate into undetected violations.

We already cite Liu et al. as closely related work and will further clarify these distinctions in the revised paper.

Our response continues in the following comment (Q2–Q4).

Continuation of our response (Q2-Q4).

Q2: Why were additional gradient-based and intervention-based attribution methods not included?

Our evaluation already includes multiple complementary attribution baselines, contrary to the suggestion that only one alternative attribution method was considered. In particular, GIF− measures realized-token logit-gradient sensitivity using the same model gradients, directly isolating the contribution of GIF's distribution-aware geometry. We additionally evaluate RTBAS-LLA, full RTBAS in computationally feasible settings (Appendix F.1), BM25, random ranking, and GPT-5.5 xhigh attribution (Appendix F.2). The suggested methods also introduce substantial computational overhead at our evaluation scale. Integrated Gradients (Sundararajan et al., 2017) requires multiple backward passes per analyzed token, typically tens or more to approximate the path integral, compared with a single backward pass for our optimized GIF implementation. Representation erasure (Li et al., 2016) similarly requires repeated forward evaluations under input interventions. Both introduce additional methodological choices concerning attribution targets, reference inputs, or interventions.

Our evaluation therefore prioritizes computationally scalable, directly relevant baselines while including a controlled gradient-based ablation that isolates GIF's central methodological contribution. We believe these comparisons adequately support the conclusions of our evaluation.

Q3: Can GIF enforce confidentiality and integrity policies without an LLM-based judge?

Yes. GIF's information-flow measurement is entirely independent of the downstream policy decision procedure. Given security labels and a policy specifying prohibited flows, GIF can supply the signal for direct threshold-based enforcement without an LLM judge.

We employ an LLM-based declassifier because information flow alone does not determine whether a dependency violates policy. For example, untrusted information may legitimately influence an agent's actions, while disclosure of confidential information may be explicitly authorized by the user. The declassifier distinguishes permitted from prohibited dependencies using semantic context.

Importantly, Section 4.1 evaluates GIF's policy-relevant attribution signal independently of any LLM judge. Using benchmark-defined policy-violation labels, we measure how well GIF's source rankings distinguish violating from non-violating trajectories through PSP@k, AUROC, and AUPRC. GIF outperforms attention-based attribution across all three benchmarks on average and generally improves over the realized-token gradient baseline. These results directly demonstrate that GIF provides a useful policy-relevant signal independently of downstream semantic declassification.

This separation between information-flow tracking and declassification follows established IFC principles and is explicitly described in Appendix D.3. The LLM judge is therefore an optional policy-decision component, not a requirement of GIF itself.

Q4: What does the Gaussian-channel construction contribute, and what would be lost without it?

The Gaussian channel is not necessary to derive the Fisher pullback or its local KL expansion. Its purpose is to provide an explicit information-theoretic semantics for the geometric quantity, which is particularly important for security applications requiring sound information-flow analysis.

By constructing a stochastic channel whose local KL geometry matches that of the autoregressive model, we obtain a closed-form mutual-information measure rather than treating the Fisher trace solely as a sensitivity or attribution score.

Crucially, our local soundness theorem establishes:

I(UT;Y)≤CapT(τ;x)+o(τ).I(U_T;Y)\leq\mathrm{Cap}_T(\tau;x)+o(\tau).

Without this construction, the Fisher trace would retain its local sensitivity interpretation, and a first-order information-flow bound could still be derived directly. However, we would lose the explicit surrogate-channel mutual-information interpretation and the corresponding capacity-based formulation of our soundness guarantee. This is precisely the distinction between using information geometry for attribution and establishing a quantitative information-flow framework.

We are happy to answer any remaining questions in the discussion period. If our rebuttal has addressed your concerns, we would appreciate a score update.

Reviewer_Fos2OpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Geometric Information Flow (GIF), a framework that uses the LLM’s Jacobian and local output geometry to measure how input spans influence model outputs. It establishes a local soundness guarantee and enables scalable estimation through automatic differentiation and randomized trace estimation. Experiments on AgentDojo, MSB, and AgentDAM evaluate GIF for source localization and evidence selection for LLM-based policy decisions, reporting reduced downstream token usage and transfer to smaller surrogate models.

Critical strengths
  1. Well-motivated mathematical interpretation of a distribution-sensitive attribution score. Equations (1)-(5) connect the Jacobian, softmax Fisher geometry, local KL curvature, and Gaussian-channel mutual information. This explains what the score measures more precisely than an attention ranking or sensitivity of only the realized token's logit. The paper also explicitly acknowledges that the result is local and that zero first-order influence does not imply independence of the nonlinear channel (Appendix B.3.3). This conceptual contribution is valuable even without a full security guarantee.

  2. Broad experiments show that GIF helps select relevant evidence for LLM-based safety checks. The experiments span both integrity and confidentiality. Across six models, Table 1 shows stronger average localization than final-layer attention on all three benchmarks. Low-rank estimates preserve rankings well. These results make the method a promising component for analyzing and screening agent actions.

Critical weaknesses
  1. Cost reporting is insufficient to assess the practical efficiency advantage. The paper lack comparing GIF’s attribution cost and runtime with the baselines in Table 1. Figure 2 reports downstream judge tokens without quantifying GIF computation. The greater-than-81x claim also changes the judge model, so the savings cannot be attributed solely to evidence selection. Comparisons on the same trajectories should report both attribution overhead and total attribution-plus-judging cost at comparable detection quality.

  2. The evaluation does not establish the safety-utility tradeoff under actual enforcement. Section 4.2 and Appendix D.3 describe an evaluation-time procedure that classifies a trajectory as violating policy if any candidate action receives DENY. The reported detection metrics are appropriate for evaluating this component, but the manuscript does not establish that these decisions were enforced during agent execution or report the resulting task-completion rates. It is therefore unclear how false denials, termination, or replanning would affect legitimate tasks.

Questions for authors
  1. How does GIF’s attribution cost and runtime compare with the Table 1 baselines on the same trajectories? Please clarify the configurations underlying the 81× claim and report total attribution-plus-judging cost, including GIF computation, at comparable detection quality.

  2. Were DENY decisions enforced during agent execution or used only to classify recorded trajectories? If enforced, how were denied actions handled, and what were the resulting task-completion and attack-success/leakage rates compared with execution without the defense?

2 comments

We thank the reviewer for their constructive feedback and recognition of GIF's theoretical and empirical contributions. We address both concerns below.

Q1: Computational overhead and the 81× cost comparison

GIF's computational overhead is modest, requiring a single backward pass per analyzed token through the model being analyzed. Only a small number of output tokens need to be analyzed, corresponding to security-relevant actions such as generated tool-call names and parameters. Importantly, our surrogate-transfer results (Figure 3) demonstrate that the analysis model need not be the trajectory-generating model: a 0.6B surrogate preserves downstream detection performance even for models up to 200× larger. This enables efficient information flow analysis independently of the target model's size.

We emphasize that the 81x figure explicitly refers to downstream model inference cost, not total pipeline cost, as stated in Section 4.2. The comparison demonstrates that GIF enables a substantially smaller declassifier (Qwen3-8B) to achieve detection performance comparable to GPT-5.5 xhigh, rather than claiming an 81× end-to-end cost reduction or attributing the entire saving to evidence selection.

Q2: Offline evaluation versus live enforcement

Our evaluation deliberately operates on recorded trajectories to isolate information-flow analysis and downstream policy-decision quality, rather than evaluating a complete agent enforcement system. The declassifier is an evaluation-time decision procedure; DENY decisions were not enforced during agent execution.

Evaluating methods on fixed trajectories ensures that comparisons are made against identical agent behavior and ground-truth policy violations, without introducing confounding variation from stochastic agent execution. Live enforcement would address a different research question: how agents respond to denied actions through termination, replanning, or recovery. These behaviors depend on the enforcement architecture and are orthogonal to GIF's information-flow semantics.

Accordingly, we do not claim to establish the safety-utility tradeoff of a deployed enforcement system. Rather, our experiments establish that GIF provides effective information-flow analysis and supports accurate policy-violation detection, which are the objectives of this work.

We are happy to answer any remaining questions in the discussion period. If our rebuttal has addressed your concerns, we would appreciate a score update.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.