acceptodds
Under review as a conference paper at ICLR 2027

MemSAGE: Asymmetric Memory Beamforming for Contextual Integrity in LLM Memory

Abstract

Modern Large Language Models (LLMs) are evolving to incorporate persistent memory, enabling them to retain user information across interactions and thereby improve user experience and task performance. However, this capability can be a double-edged sword. Recent studies have revealed the contextual integrity problem of persistent memory, where LLMs utilize retained information indiscriminately without assessing its contextual appropriateness, resulting in unwarranted disclosure of sensitive information. Existing approaches to this problem either regulate memory use externally via explicit prompting and rigid retrieval control, or internalize contextual reasoning into LLMs, leaving unmet the asymmetric, context-specific demands on retained memories. To address this gap, we propose MemSAGE, a novel plug-in framework for enforcing the contextual integrity of LLM MEMory via Stiefel-based Asymmetric Geometric bEamforming. MemSAGE presents a unique twist: it trains a lightweight hypernetwork to dynamically generate Stiefel frames for each context and the retained memories, forming a nested permission–demand geometry that directs memory use during generation. To operationalize this structure, MemSAGE performs asymmetric memory beamforming over LLM attention, suppressing context-inappropriate memories while directionally amplifying obligatory ones through geometry-aligned query–key interactions. Notably, MemSAGE is applied at inference time without altering the retained memories or updating the base model parameters. Extensive experiments across various LLM backbones demonstrate that MemSAGE significantly enhances the contextual integrity of LLM memory compared with existing baselines, simultaneously reducing memory violations and improving coverage.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Contextual Integrity, LLM Memory, Inference-time Steering

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_vLu2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MemSAGE addresses contextual-integrity violations in persistent-memory LLMs by learning context-conditioned permission and demand subspaces on a Stiefel manifold. It derives asymmetric geometric scores for each retained memory, then modifies attention to suppress norm-violating memories and amplify task-obligatory ones without updating the backbone. Experiments on CIMemories and CI-RL with four LLMs report lower disclosure violations, improved completeness.

Critical strengths
  1. The paper introduces a technically distinctive formulation: nested permission-demand Stiefel geometry is converted into asymmetric, memory-specific attention control, providing a more granular alternative to prompt constraints or hard retrieval filtering.
  2. The three-stage design is coherent and modular: contextual framing, spectral beamforming, and token-level steering connect an explicit contextual-integrity model to a frozen-backbone inference intervention.
Critical weaknesses
  1. The rationale for using the Stiefel manifold is insufficiently developed. The paper explains the benefits of orthonormal frames, but does not clarify why the present problem specifically requires a Stiefel-manifold formulation. It should discuss plausible alternatives and explain why Stiefel geometry is preferable to other orthogonalization or subspace-learning approaches.
  2. The performance of the permission and demand classifiers is not reported. Equation 6 directly constructs supervision from MviolM_{viol} and MdemM_{{dem}}, yet the paper provides no classification accuracy, recall, or confusion matrices for these two prediction tasks.
  3. Table 5 does not establish the necessity of the Stiefel geometry. It only replaces the spectral gate with MLP and bilinear readouts, while retaining the same Stiefel-frame construction, training objectives, and attention-steering pipeline. Thus, it shows only that the spectral readout outperforms two simpler scoring heads, not that orthogonality, principal-angle information, or the Stiefel manifold is necessary.
  4. The construction of the training trajectories used in Equation 7 is unclear. Computing Ai=Falign(Q,Pc,midem,Ki)A_i=F_{\mathrm{align}}(Q,P_{c,m_i}^{\mathrm{dem}},K_i) appears to require query representations from the complete generation trajectory. The paper does not specify which model and steering configuration generate these trajectories, or whether reference responses or future generation states are used during training.
Questions for authors

See weakness

6 comments
AuthorsOpenReview ↗

Response to Reviewer vLu2 (1/2)

Thank you for your careful assessment and constructive suggestions. We provide the following clarifications to address your concerns:

1. Regarding the rationale and empirical justification for Stiefel geometry (W1, W3):

We appreciate your insightful feedback. In response, we provide the following clarifications and supporting evidence:

(1) Design rationale for Stiefel geometry: We would like to clarify that the Stiefel formulation is essential to the spectral alignment, nested geometry, and memory steering of MemSAGE. Specifically:

  • Scale-independent subspace comparison: With orthonormal frames, the singular values of His\mathbf{H}_i^s in Equation 3 correspond to the cosines of principal angles between subspaces. This ensures that the spectral affinities reflect directional alignment without being confounded by arbitrary scaling of the basis vectors.
  • Structurally guaranteed geometric nesting: The nested Stiefel construction places the demand subspace within the permission subspace and guarantees that its spectral affinities never exceed the corresponding permission affinities. This structurally mirrors the containment of task obligation within permissible disclosure, ensuring that demand alignment requires at least equally strong permission alignment.
  • Bounded transformation of memory keys: The Stiefel formulation ensures that the demand-conditioned operator, constructed as a composition of orthogonal projections, does not increase key norms. This prevents the geometric transformation from introducing arbitrary scale amplification, keeping the enhancement in Equation 10 dependent on directional query–key compatibility and the explicit steering coefficients.

(2) Controlled evaluation of the Stiefel geometry: To directly examine the necessity and effectiveness of the Stiefel formulation, we conduct complementary experiments under the following setup:

Setup: We compare MemSAGE with two variants: unnormalized orthogonalization, which retains mutually orthogonal columns while allowing their norms to vary, and unconstrained subspace construction, which directly uses the hypernetwork outputs without orthogonalization or normalization. Each variant applies the same modification to the context and memory frames and the factor constructing the nested demand frame. All other training and steering settings remain unchanged.

Table R1. Comparison of alternative geometric constructions.
MethodViolation@5 (%) ↓Completeness (%) ↑
MemSAGE (Stiefel geometry)44.0147.31
Unnormalized orthogonalization38.1623.45
Unconstrained subspace construction47.5230.38

We highlight two observations:

  • Normalization preserves context-dependent memory control. Removing normalization reduces Violation@5 from 44.01% to 38.16%, but substantially lowers Completeness from 47.31% to 23.45%. Without normalization, the spectral affinities depend on basis-vector norms as well as subspace alignment. This allows changes in memory-specific scale to substitute for contextual alignment, potentially weakening MemSAGE’s ability to distinguish when the same information should be suppressed or retained.
  • Orthogonality preserves distinct directions for memory control. Unconstrained subspace construction increases Violation@5 from 44.01% to 47.52% while reducing Completeness from 47.31% to 30.38%. Without orthogonality constraints, frame columns can become redundant, reducing the effective dimensionality of the permission–demand representation. Such redundancy also distorts the spectral affinities used to assess contextual alignment, weakening MemSAGE’s ability to separate appropriate memory use from inappropriate disclosure.

We believe these clarifications and experimental results provide a more comprehensive understanding of the rationale for and effectiveness of the Stiefel formulation in MemSAGE. We will incorporate these discussions and additional results into the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer vLu2 (2/2)

2. Regarding the performance of the permission and demand classifiers (W2):

We appreciate your feedback on this aspect. We evaluate the permission and demand predictions against the ground-truth contextual labels using AUROC, accuracy, recall, and precision, with results reported in Table R2. For recall and precision, the positive class is norm-violating memories for permission prediction and task-obligatory memories for demand prediction.

Table R2. Standalone performance of the permission and demand classifiers.
ClassifierAUROCAccuracy (%)Recall (%)Precision (%)
Permission0.81175.581.780.9
Demand0.88080.676.383.5

We highlight three observations:

  • Permission scores effectively identify norm-violating memories. The permission classifier achieves a recall of 81.7%, indicating the ability of the learned permission scores to identify memories that warrant suppression. This is critical for contextual integrity since missed norm-violating memories may lead to inappropriate disclosure during generation.
  • Demand scores reliably identify task-obligatory memories. The demand classifier achieves an accuracy of 80.6% and a precision of 83.5%, supporting selective recognition of task requirements. This is important for demand-aware steering, as mistaking unnecessary information for task-obligatory content can divert emphasis from information genuinely required for task completion.
  • Permission and demand scores provide graded discrimination between memory categories. Their AUROC values of 0.811 and 0.880 indicate that both scores meaningfully rank memories according to their contextual labels. This extends their utility beyond binary decisions at a particular threshold, supporting the permission–demand geometry underlying MemSAGE’s memory steering.

We believe these results provide a more comprehensive understanding of the effectiveness of MemSAGE’s permission and demand predictions. We will incorporate these results and the corresponding analysis into the revised manuscript.

3. Regarding the construction of training trajectories (W4):

Thank you for your feedback. We provide the clarifications as follows:

(1) Trajectory generation: We generate the reference responses offline using the same frozen backbone, without applying MemSAGE steering. For each training context, we remove norm-violating memories from the generation prompt and explicitly instruct the model to incorporate task-obligatory memories based on the ground-truth labels. These responses are generated once using only the training profiles and remain fixed throughout training.

(2) Query extraction and alignment supervision: We append each reference response to the original complete prompt containing all memory entries and perform a teacher-forced forward pass through the frozen backbone. We extract queries from the response positions and keys from the memory spans to compute the alignment objective in Equation 7. The standard causal mask prevents each position from attending to subsequent response tokens. These response queries serve only as training supervision for the demand-conditioned projection. During inference, the permission scores and projections are computed from the context and memory representations without access to reference responses or future generation states.

We will clarify the reference-response construction and alignment supervision in the revised manuscript.

Dear auhtor,

Thank you for the response. However, my main concerns are not addressed:

  1. W1: The rebuttal only restates generic benefits of orthonormal frames, which I had already acknowledged. It still does not justify why the Stiefel manifold is required over standard alternatives such as Grassmann parameterization, QR/Cayley orthogonalization, or soft orthogonality constraints. Neither the paper nor the rebuttal compares against any of them. The core design choice remains unjustified in both sufficiency and necessity.
  2. W3: Table R1 does not support the authors’ claim. The variants are clearly degenerate: the unnormalized one reaches only 23.45% Completeness, worse than Vanilla (37.65%, Qwen3-8B in Table 1). At most, the results show that the pipeline was built around orthonormality. They do not show that Stiefel geometry is necessary.
  3. Revision: The authors promise revisions but have not uploaded a revised PDF, so none of the claimed changes can be verified.

Given these unresolved issues and the unrevised manuscript, I retain my original score of 2.

Reviewer_D7w1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses contextual integrity (CI) in memory-augmented LLMs, where information stored in persistent memory may surface in contexts whose norms forbid its disclosure. The authors propose MemSAGE, which performs inference-time, attention-level memory steering without modifying the backbone LLM. Specifically, MemSAGE trains hypernetworks to generate Stiefel frames for each context and user memory, and uses the spectral alignment to derive beamforming gains that decide whether each memory should be suppressed or amplified. The experiments show that MemSAGE reduces norm-violating information disclosures while improving the coverage of task-obligatory memories.

Critical strengths
  1. The nested geometry and geometric alignment offer a well-motivated fit for modeling CI disclosure norms, and this design choice is empirically justified. This is consequential to my overall assessment because one of the the paper’s main contributions lies in its geometric formulation and its encoding of the asymmetric structure of CI norms.

  2. MemSAGE remains effective compared to baselines that update the backbone LLM, while keeping the backbone LLM frozen. Across four backbone LLMs and two benchmarks, it generally achieves lower violation rates than the baselines, with comparable or better completeness.

Critical weaknesses
  1. It remains unclear whether MemSAGE’s violation reduction stem from permission-based suppression, as claimed, or from other mechanisms such as amplification. The paper attributes the higher Violation@5 of hard filtering to the intervention locus, but this explanation remains qualitative. Since filtered memories cannot be disclosed, the remaining violations must come from memories deemed permissible, and the paper does not verify which mechanism reduces violations from these memories under MemSAGE.

  2. The robustness and generalization of the improvements are difficult to assess. On CIMemories, the in-distribution benchmark, evaluation relies on a single random split with only three held-out profiles, and no variance across splits or runs is reported; profile-level cross-validation (e.g., leave-one-profile-out or k-fold) would make these results more reliable. On the CI-RL test set, the out-of-domain benchmark, MemSAGE still attains the lowest violation rates, but its completeness is comparable to or lower than that of the strongest baselines for several backbones, suggesting that the claimed improvement in the coverage of task-obligatory memories holds primarily in-distribution.

Questions for authors
  1. Since memories removed by hard filtering cannot be disclosed, the remaining violations in Table 3 must come from memories the gate deems permissible. The paper attributes MemSAGE’s further reduction to the intervention locus; could the authors provide direct evidence for this, such as the attention mass assigned to these memories under each method? What is the standalone precision and recall of the permission scores?

  2. Could the authors report CIMemories results under profile-level cross-validation or multiple random splits, together with variance across runs?

  3. Why do the completeness gains diminish on the CI-RL test set? Is this attributable to structural differences between the benchmarks or to the higher completeness ceiling?

4 comments
AuthorsOpenReview ↗

Response to Reviewer D7w1 (1/2)

We appreciate your thoughtful comments and recognition of our work. We address your concerns as follows:

1. Regarding the mechanisms underlying violation reduction beyond hard filtering (W1, Q1):

We appreciate your insightful feedback on this aspect. In response, we provide the following analyses to further examine the role of permission-based suppression and the performance of the permission scores:

(1) Attention allocation to retained memories: Following your suggestion, we compare attention allocation under hard filtering, suppression-only MemSAGE, and full MemSAGE on Qwen3-8B. We report the per-memory attention mass relative to the vanilla model. The retained norm-violating memories are grouped by their layer-averaged permission scores into those near the threshold (0.5-0.7) and those above 0.7.

Table R1. Per-memory attention mass relative to vanilla.
Memory categoryHard filteringSuppression onlyFull MemSAGE
Norm-violating, below threshold–0.35×0.36×
Norm-violating, retained near threshold1.74×1.12×0.98×
Norm-violating, retained above 0.72.01×1.51×1.68×
Task-obligatory1.66×1.30×2.27×

We highlight two observations:

  • Permission-based suppression limits the attention to the retained norm-violating memories compared with hard filtering. For retained norm-violating memories, hard filtering raises their attention to 1.74× and 2.01× the vanilla level, since removing memories from the input redistributes attention to the retained ones. Suppression reduces these values to 1.12× and 1.51×, providing direct evidence that permission-based suppression regulates these memories even when they pass the filtering threshold.
  • Amplification mainly promotes task-obligatory memory use while complementing suppression. It raises the attention to task-obligatory memories from 1.30× to 2.27× the vanilla level. By redirecting attention, it further lowers the attention to norm-violating memories near the threshold. These results support the complementary roles of the two mechanisms in prioritizing obligatory memories while limiting attention to retained norm-violating memories.

(2) Standalone performance of the permission and demand scores: We additionally evaluate the permission and demand predictions against the ground-truth contextual labels using AUROC, accuracy, recall, and precision, with results reported in Table R2. For recall and precision, the positive class is norm-violating memories for permission prediction and task-obligatory memories for demand prediction.

Table R2. Standalone performance of the permission and demand classifiers.
ClassifierAUROCAccuracy (%)Recall (%)Precision (%)
Permission0.81175.581.780.9
Demand0.88080.676.383.5

We highlight three observations:

  • Permission scores effectively identify norm-violating memories. The permission classifier achieves a recall of 81.7%, indicating the ability of the learned permission scores to identify memories that warrant suppression. This is critical for contextual integrity since missed norm-violating memories may lead to inappropriate disclosure during generation.
  • Demand scores reliably identify task-obligatory memories. The demand classifier achieves an accuracy of 80.6% and a precision of 83.5%, supporting selective recognition of task requirements. This is important for demand-aware steering, as mistaking unnecessary information for task-obligatory content can divert emphasis from information genuinely required for task completion.
  • Permission and demand scores provide graded discrimination between memory categories. Their AUROC values of 0.811 and 0.880 indicate that both scores meaningfully rank memories according to their contextual labels. This extends their utility beyond binary decisions at a particular threshold, supporting the permission–demand geometry underlying MemSAGE’s memory steering.

We will incorporate these additional analyses and clarifications into the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer D7w1 (2/2)

2. Regarding the robustness of the CIMemories evaluation (W2, Q2):

Thank you for the suggestion. We conduct additional experiments comparing MemSAGE with SelfCI across three profile-level splits on CIMemories using Qwen3-8B. Each split uses seven profiles for training and three for testing, with all other settings unchanged. Table R3 reports the mean and standard deviation across splits.

Table R3. Performance across three profile-level splits.
MethodViolation@5 ↓Completeness ↑
SelfCI54.10 ± 1.6443.86 ± 2.51
MemSAGE44.99 ± 1.4847.12 ± 1.57

MemSAGE consistently achieves lower Violation@5 and higher Completeness across all three splits, reducing average Violation@5 from 54.10% to 44.99% while yielding higher Completeness (47.12% vs. 43.86%). These results support that MemSAGE’s gains are robust across different profile splits.

We will incorporate these additional experimental results into the revised manuscript.

3. Regarding the completeness gains on the out-of-domain benchmark (W2, Q3):

We appreciate your feedback. We clarify that MemSAGE’s performance on the CI-RL test set may be attributed to structural differences between the two benchmarks. Specifically, CIMemories contains approximately 150 memories per profile. Task-obligatory memories compete with a much larger pool of contextual information during generation, making demand-aware amplification particularly beneficial in this setting. In contrast, each CI-RL scenario involves far fewer memories, making the required information easier to utilize. This is reflected in the vanilla model’s Completeness, which reaches 84.07%–88.14% on the CI-RL test set, compared with 37.15%–45.97% on CIMemories, leaving less room for additional gains from amplification.

We will include this analysis in our revised manuscript.

Thank you for the detailed response and additional experiments. I appreciate the authors' efforts to address my concerns.

The attention analysis and standalone classifier results provide useful evidence for the roles of suppression and amplification. The additional profile-level split experiments also strengthen confidence in the robustness of the reported improvements.

I will maintain my positive rating.

Dear Reviewer,

Thank you again for your time and continued positive assessment. We are glad our rebuttal addressed your concerns.

Please feel free to raise any further questions during the remaining discussion period.

Best regards,

Authors

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

MemSAGE addresses contextual integrity in memory-augmented LLMs by controlling which memories influence generation. It learns Stiefel representations for contexts and memories, with a demand subspace nested within a permission subspace. These representations guide two complementary attention interventions: suppressing contextually inappropriate memories and amplifying information required by the task. Only auxiliary components are trained; the underlying LLM remains frozen. The evaluation covers four LLM backbones on CIMemories and an out-of-domain CI-RL test set. The paper reports reductions in disclosure violations while generally maintaining or improving completeness. Additional experiments examine individual components, hard filtering, oracle supervision, and alternatives to geometric scoring.

Critical strengths
  1. The method captures an important distinction between suppressing information and actively using it.

  2. The empirical evaluation includes relevant and informative controls. The method is tested across four backbones and a second dataset.

Critical weaknesses
  1. The robustness and comparability of the gains are not fully established. The CIMemories evaluation uses one split containing only three held-out profiles and 147 contexts, with no reported variation across training seeds or alternative splits.

  2. The mechanistic claims should be distinguished from empirical leakage reduction. Reducing attention to a memory’s original tokens does not necessarily prevent its information from propagating through other contextualized tokens or cached representations.

Questions for authors
  1. Were all trainable baselines adapted on the same CIMemories training profiles, with CI-RL excluded from training and tuning for every method? Please clarify training budgets, hyperparameter selection, and overlap in task templates or recipients across profile splits. How stable are the results across seeds or alternative splits?

  2. What accounts for the remaining 11.44% violation rate under oracle asymmetric control? Does oracle hard masking remove information before prefill, or only mask attention to the original memory spans? Distinguishing these interventions would help explain the observed gap.

  3. How are context and memory representations extracted and pooled, and how are the resulting frames used across attention heads? Is steering applied during both prefill and decoding? Please also clarify how the demand score from Equation 5 is used, since it does not explicitly appear in Equation 10, particularly when constructing the oracle variant.

  4. Can the methods be compared at matched completeness or matched violation rates by varying the suppression threshold and amplification strength? This would clarify whether MemSAGE offers a consistently better trade-off rather than a favorable single configuration.

  5. Report auxiliary parameter counts, training cost, prefill latency, decoding throughput, and additional memory use. These measurements are needed to assess the practical implications of intervening at every attention layer and head.

3 comments
AuthorsOpenReview ↗

Response to Reviewer kocS (1/3)

Thank you for your time and constructive feedback. We provide the following clarifications and responses to address your concerns:

1. Regarding the robustness and comparability of the gains (W1, Q1):

We appreciate your feedback on the robustness and comparability of our evaluation. In response, we provide the following clarifications and additional results:

(1) Comparison settings and data splitting: We would like to clarify that all trainable baselines are trained on the same seven CIMemories profiles, with the remaining three profiles held out for evaluation. We adapt CI-RL using its released implementation on our CIMemories training split. The CI-RL benchmark is excluded from training and hyperparameter tuning for all methods.

Regarding overlap across splits, all CIMemories profiles share the same 49 task–recipient combinations. Thus, the evaluation assesses generalization to unseen users and their memories, rather than unseen task–recipient combinations. We follow the hyperparameter settings and training schedules reported in the baselines’ released implementations when adapting them to the CIMemories training split.

(2) Stability across profile splits: We conduct additional experiments comparing MemSAGE with SelfCI across three profile-level splits on CIMemories using Qwen3-8B. Each split uses seven profiles for training and three for testing, with all other settings unchanged. Table R1 reports the mean and standard deviation across splits.

Table R1. Performance across three profile-level splits.
MethodViolation@5 ↓Completeness ↑
SelfCI54.10 ± 1.6443.86 ± 2.51
MemSAGE44.99 ± 1.4847.12 ± 1.57

MemSAGE consistently achieves lower Violation@5 and higher Completeness across all three splits, reducing average Violation@5 from 54.10% to 44.99% while yielding higher Completeness (47.12% vs. 43.86%). These results support that MemSAGE’s gains are robust across different profile splits.

We will incorporate these clarifications and additional experimental results into the revised manuscript.

2. Regarding the mechanistic claims and remaining violations under oracle control (W2, Q2):

We appreciate your insightful feedback on this aspect. In response, we provide the following clarifications:

(1) For the attention-level intervention and empirical leakage reduction: We acknowledge that information associated with a memory may remain accessible through contextualized representations, even when attention to its original tokens is reduced. However, we would like to highlight that prior studies [1, 2] have demonstrated that targeted attention interventions can influence how LLMs use contextual information and shape their generation behavior. MemSAGE is thus designed to regulate memory participation at the attention interface through permission-based suppression and demand-aware amplification, thereby guiding generation toward contextually appropriate memory use.

Our experiments provide complementary evidence for this design: Figure 4 shows how MemSAGE changes attention allocation across memory categories, while the Violation@5 results demonstrate reduced inappropriate disclosure in generated outputs. These findings support the effectiveness of our memory-steering mechanism, without implying complete isolation of memory information.

(2) For the explanation of the oracle analysis: We clarify the two oracle variants as follows:

  • Oracle hard masking indeed masks attention to the original spans of norm-violating memories without removing them from the input before prefill, with no demand-side amplification applied.
  • Oracle asymmetric control instead uses ground-truth labels to strongly suppress private memories and amplify task-obligatory ones, while retaining the learned demand projection.

The remaining violation rate under oracle asymmetric control arises from the unintended deduction of norm-violating information when elaborating on task-obligatory memories. We note that oracle control operates on memory entries, whereas the judge assesses the facts disclosed in generated responses. MemSAGE’s design of demand-side amplification promotes the utilization of task-obligatory memories, leading to more detailed responses in which the model may inadvertently deduce and disclose associated private facts. For example, mentioning a task-obligatory memory that someone is on sick leave may allow the model to infer an underlying health issue, even when the private memory describing their health condition is masked.

We will incorporate these clarifications into the revised manuscript.

Reference

[1] Zhang et al. Tell Your Model Where to Attend: Post-hoc Attention Steering for LLMs. ICLR, 2024.

[2] Yan et al. Don't Take Things Out of Context: Attention Intervention for Enhancing Chain-of-Thought Reasoning in Large Language Models. ICLR, 2025.

AuthorsOpenReview ↗

Response to Reviewer kocS (2/3)

3. Regarding representation extraction and the implementation of memory steering (Q3):

Thank you for your feedback. We provide the following answers to each of your questions:

(1) For context and memory representation extraction: We obtain context and memory representations from the input hidden states of the self-attention module at every layer of the frozen backbone. We apply mean-pooling over the context span and each memory’s token span. The resulting context and memory representations, concatenated with a learnable layer embedding, are passed to the corresponding hypernetworks to construct the context and memory frames in Equation 2.

(2) For frame usage across attention heads: The frames are shared across attention heads within each layer. Their encoding dimension matches the attention head dimension. The permission score is a single scalar per memory and per layer, and the resulting suppression bias is applied identically across all heads. Within each head, the demand-conditioned projection operates on the corresponding memory keys, whose compatibility with that head’s current query determines the amplification in Equation 10.

(3) For steering during prefill and decoding: Memory steering is applied during both prefill and decoding. Specifically, during prefill, the suppression bias is added at every layer to the attention logits from prompt positions to memory tokens, subject to the causal mask. The permission scores and projected memory keys are computed once at this stage and cached. During decoding, the cached suppression bias is applied to each newly generated token's attention to memory tokens, together with the demand-aware amplification, which is applied only during decoding. The original memory tokens remain in the input throughout this process.

(4) For the role of the demand score: We would like to clarify that the demand score in Equation 5 is used only in the training loss, which shapes the nested demand geometry under ground-truth demand supervision. During generation, this geometry is used to construct the demand-conditioned projection in Equation 10, allowing amplification to depend on the current query and individual memory keys. Thus, the scalar demand score is not directly used during inference. In the oracle variant, ground-truth demand labels determine which memories receive amplification, while the learned demand projection determines the directional adjustment.

We will incorporate these implementation details into the revised manuscript.

4. Regarding comparisons at matched completeness or violation rates (Q4):

Thank you for the suggestion. We conduct complementary experiments to compare MemSAGE with the baselines at comparable violation or completeness levels.

(1) Comparison at matched violation rates: We vary the suppression threshold τ\tau on Qwen3-8B using CIMemories, with results reported in Table R2. At τ=0.4\tau=0.4, MemSAGE achieves a Violation@5 of 53.42%, comparable to SelfCI’s 54.76%, while yielding higher Completeness (47.80% vs. 44.13%). This improvement in Completeness demonstrates that MemSAGE can preserve more task-obligatory information under comparable disclosure control.

Table R2. Effect of the suppression threshold on Qwen3-8B evaluated on CIMemories. Metrics are reported as percentages.
Suppression threshold τViolation@5 ↓Completeness ↑
0.360.4648.38
0.453.4247.80
0.544.0147.31
0.643.2645.99
0.739.5541.41

(2) Comparison at matched completeness: We vary the amplification strength κ\kappa under the same setting, with results reported in Table R3. At κ=0.6\kappa=0.6, MemSAGE achieves a Completeness of 44.55%, comparable to SelfCI’s 44.13%, while yielding lower Violation@5 (52.40% vs. 54.76%). This demonstrates that MemSAGE provides stronger disclosure control while maintaining comparable coverage of task-obligatory information.

Table R3. Effect of the amplification strength on Qwen3-8B evaluated on CIMemories. Metrics are reported as percentages.
Amplification strength κViolation@5 ↓Completeness ↑
0.456.1043.20
0.652.4044.55
0.849.6546.70
1.044.0147.31
1.247.8549.61

Discussion: These results indicate that MemSAGE’s advantage holds at comparable operating points rather than only at a favorable single configuration, demonstrating a better privacy–utility trade-off. We will incorporate these additional results and the corresponding discussion into the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer kocS (3/3)

5. Regarding parameter counts and computational overhead (Q5):

Thank you for raising this point. Following your suggestion, we quantify the parameter, training, and inference overhead of MemSAGE across all four backbones, with results reported in Table R4. Inference measurements are repeated three times on the CIMemories test contexts, with variation below 0.5%.

Table R4. Computational overhead across four backbones. Parentheses in the last three columns report changes relative to vanilla. Training times are reported as estimated upper bounds.
BackboneMethodAux. params (% of backbone)Training time (h)Prefill latency (s) ↓Decoding throughput (tok/s) ↑Peak GPU memory (GiB) ↓
Mistral-7BVanilla––0.1276.314.11
MemSAGE9.39M (0.13%)≤ 2.10.89 (+0.77)67.7 (88.7%)14.28 (+1.2%)
Qwen3-8BVanilla––0.1266.615.81
MemSAGE9.39M (0.11%)≤ 2.10.98 (+0.86)59.6 (89.5%)15.98 (+1.1%)
Qwen3-32BVanilla––0.4519.662.00
MemSAGE10.44M (0.03%)≤ 6.31.98 (+1.53)18.5 (94.4%)62.30 (+0.5%)
Llama-3.3-70BVanilla––0.899.70132.97
MemSAGE13.59M (0.02%)≤ 7.23.00 (+2.11)9.36 (96.5%)133.32 (+0.3%)

We highlight three observations:

  • Small parameter and memory overhead. MemSAGE introduces auxiliary parameters equivalent to only 0.02%–0.13% of the backbone and increases peak GPU memory by 0.3%–1.2%. These results demonstrate that sharing hypernetwork parameters across layers and reusing frames across heads within each layer keeps the parameter overhead small, despite steering at every attention layer and head.
  • Preserved decoding throughput. MemSAGE retains 88.7%–96.5% of vanilla decoding throughput. Permission scores and projected memory keys are computed during prefill and cached for reuse, avoiding their recomputation at each decoding step.
  • Additional computation during prefill. MemSAGE adds 0.77–2.11 seconds of prefill latency to construct the frames and prepare the steering quantities. Note that this is a one-time cost and does not grow with the number of generated tokens.

We will incorporate these efficiency measurements and implementation details into the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.