acceptodds
Under review as a conference paper at ICLR 2027

BRIDGe: Behavioral Rule Installed via Dual-Gate steering

Abstract

The outcome of a language model (LM) depends as much on what it already knows as on what it is told. However, this knowledge is partly volatile: facts become outdated or contradictory in time. What we want is a conditional rule: use their memory when asked about stable facts, rely on a given context when asked about volatile facts, and defer when facts are volatile and no context is available. However no model we have tested does this, and no training installs it either: three small language model (SLM) backbones and five training objectives built for this task reach at most a third of the required behavior. Prompting does not close the gap either: instructing a model to abstain raises deferral indiscriminately. Yet fact volatility and context presence are both linearly decodable in these models, but nothing connects them to the decision. We wire them directly. We introduce BRIDGE (Behavioral Rule Installed via Dual-Gate steering), an inference time method that reads both signals from activations and steers only when both hold. Gating on context presence alone raises deferral on stable and volatile questions alike. The two-signal gate raises discrimination on every backbone, by a wide margin, leaving contextual obedience essentially unchanged. The result is an SLM that can be trusted for fresh information and a way to install rules that depend on what a model knows, something neither a prompt nor a fine-tuning objective can express.

Primary area:
foundation or frontier models, including LLMs
Keywords:
activation steering, abstention and deferral, inference-time intervention, fact volatility

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_pcE2OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies whether small language models of 7–8B scale can follow a conditional rule, answer stable facts from memory, answer volatile facts from retrieved context, and defer when a fact is volatile and no context is given. It defines “volatile” as a fact whose value changed between two EvoWiki snapshots. The authors show that prompting and five fine-tuning objectives fail to install this rule. Through probing, patching and steering, they demonstrate that volatility is linearly decodable but unused by the deferral decision. Based on this observation, the paper proposes BRIDGe, an inference-time hook that steers along the context-presence direction only when probes indicate the fact is volatile and no context is present. On three backbones, BRIDGe improves deferral discrimination over the best fine-tuned model, at a modest cost in stable retention and with contextual obedience nearly unchanged.

Critical strengths
  • The paper addresses practically relevant problem with a careful evaluation protocol. Knowing when parametric memory may be stale, as distinct from simply following context or abstaining, matters for deployed RAG system.
  • The intervention pipeline that separates volatility and presence is simple and interpretable.
  • The ablation of random and shuffled-label directions do not raise DD supports the design within paper's own setting.
  • The released code aids reproducibility.
Critical weaknesses
  • The central mechanistic claims, and the framing of volatility as “the model’s own estimate,” are not sufficiently supported. Volatility is defined by whether a value changed between two arbitrary snapshots. It is unclear how a model could know which facts will change beyond correlates such as relation type or entity category. If so, the volatility probe may detect question type rather than an internal belief about staleness. The retail experiment on Section 6 states that near-ceiling deferral there comes from “a small, learnable tag set,” which reinforces this concern.
  • The strength of the diagnosis is not clear enough. The causal conclusion rests largely on activation patching, however but the patching uses only 7 valid volatile/stable pairs on Llama, and the steering comparison of \hat_v versus \hat_c is run only on Llama. Also in Mistral, the ablation has no significant effect on either split.
  • Conditional activation steering is well studied, and closely related methods are cited, e.g. CAST (Lee et al., 2025), SHIFT (Li et al., 2026) , but the paper didn't compare theirs with those existing method. The baselines are mostly focused on authors’ own fine-tuning variants, a presence-only ablation of BRIDGe. Without such comparisons, it is hard to judge whether the design choices are well-founded.
  • The method contains hyperparameters of which selection criteria is unclear. Section 4.2 reports “the setting with the best trade-off,” but it is unclear whether \tau_v, \tau_c, \alpha and the layer were chosen on a split separate from the reported evaluation set. Also, Table 3, only the Llama-3.1-8B uses a manually set. Clarification is needed on which configuration produced each number.
  • Main results (Tables 2–3) come from a single run without confidence intervals.
  • In absolute terms, VRR under BRIDGe remains at 44-60%, and SDR rises roughly 3–5x (e.g., 6.1 -> 20.8 on Llama). The model still answers many volatile questions from stale memory and now declines more stable ones.
  • BRIDGe requires gold volatility labels to fit its probe. In practical setting, if such labels exist at deployment time, a classifier on the question text or a lookup may be a simpler alternative, and the paper should compare against it. Also, in the retail scenario presented as “the common production scenario,” CA-AVMD alone reaches 99.5% VRR and 94.3% CO, and BRIDGe is not applied. This weakens the case that the method is needed where the paper says it matters most.

[minor]

  • Many abbreviations are introduced quickly or before their definitions. Table 1 appears before the metrics are defined in Section 2.1. As all column names and row names are abbreviation, an illustrative figure would help greatly. Potentially, one concrete EvoWiki example per cell of the rule, showing the expected output and which metric scores it.
  • These are used in Equation 1 with only a brief introductory definition. More concrete explanation on the snapshots should be stated in the main text.
  • Specific versions are omitted in Table 2 (“Llama,” “Mistral,” “Qwen”) and named inconsistently elsewhere (“Mistral3-7B” vs. “Mistral-7B-v0.3”). Please state exact versions consistently, and indicate base or instruct.
  • Redundancy. “A deferral denoted by def(\hat{y}) = 1 if \hat{y}is an output that declines to give a value” is followed by identical definition, “def(\hat{y}) = 1 if \hat{y} is a deferral and 0 otherwise.”
Questions for authors
  • How can a model know which facts are volatile, given that volatility is defined by changes between two snapshots? Does the volatility probe generalize to held-out relation types, or mainly detect relation or question category? Please report probe accuracy and BRIDGe results by relation type and on the semi-volatile held-out split.
  • On what data were \tau_v, \tau_c, \alpha and the hook layer selected, and is it disjoint from the evaluation set behind Tables 2–3?
  • How does BRIDGe compare with an adapted conditional steering baseline such as CAST?
  • Which exact checkpoints are used (base or instruct)? Can confidence intervals or multi-seed results be provided for all the main tables, including Qwen?
  • From my understanding, the rule rewards following context (CO) even when it contradicts memory, while the introduction motivates the work by noting that retrieval can be missing or incomplete. How does BRIDGe behave with irrelevant, incomplete or incorrect context?
  • Does the representation–policy gap persist at larger scales? If volatility is linearly decodable, larger models might use it without intervention. If the paper’s scope is restricted to SLMs, this could be stated more explicitly in the title or abstract
5 comments

Thank you for the detailed and useful review. We first address the main concern around volatility and then clarify the other ones, and report the results the reviewer has asked for.

  • Volatility is a concept (facts that go stale) and whether the fact has changed between 2 timestamps is simply a way to label the data, not a definition of this concept. And as the reviewer intuited, the model is picking up on cues of the relationship type and entities. However the model would still fail to defer. And if labels are available, to distinguish volatile facts, then simpler classifiers (like tf-idf) can be good alternatives - see table B. If these cues are not seen in training, then both gates lose most of their accuracy, but the TF-IDF router ends up worse than no intervention, while BRIDGe with a logistic-probe gate fires rarely and still slightly improves (we ran this and results are in Table C). The retail case contain these cues as well, but the cues are not the bottleneck, but prior knowledge. We do not position BRIDGe as the diagnosis or the only router, but as a lightweight, interpretable intervention demonstrating how volatility-associated and context-presence features can be coupled. Our wording ("the model's own estimate") is confusing, and we will clarify it.

  • Another important aspect raised in the reviews is the causal claim: our diagnosis is based on steering while patching is used mostly for locating the signal. Presence is the strongest signal on Llama and Qwen (Mistral is comparable as well. However we agree that volatility survives ablation so we cannot measure deferral without it - we’ll add this clarification.

  • Hyperparameters alpha and layer are fixed while tau_c and tau_v were selected on the evaluation questions, which were held from fine-tuning not evaluation (we will correct it in A4). Qwen uses a larger alpha to match Llama activation magnitude. However, re-picking the parameters with 5-fold cross-validation keeps the gain significant on all three backbones (Table C). Regarding the baselines, both SHIFT and CAST are relevant related approaches. In the paper we did try conditioning on a single gate (see Tables 3 and 13) which motivates our choice of combining both context presence and volatility signals together. However, we did not directly benchmark against CAST or SHIFT and therefore make no claim of superiority over them.

  • Nonetheless, as the reviewer pointed out, there are limitations and tradeoffs. Even though BRIDGe significantly improves deferral discrimination over the fine-tuned baselines, we did not achieve 100% coverage. This remains a trade-off, and other mechanisms such as RL or stronger base representations can be further studied. We will state the 7–8B scope in the title and abstract. Further experiments could explore a relevance-aware gate: what happens when the context is also irrelevant (not only missing) as the gate right now doesn’t make that distinction.

We hope this shows that the paper’s claim (a behavioural gap that fine-tuning does not close, and a conditional intervention that narrows it) holds with the new data, and we would be grateful if the reviewer would reconsider the soundness scores.

Table A. Llama, on relation types not seen in training (higher is better).

MethodDDSR
No intervention (CA-AVMD)18.853.5
BRIDGe, probe gate22.252.0
BRIDGe (Table 3 in paper)15.639.8
Router, TF-IDF−4.716.9
Router, TF-IDF, threshold re-tuned13.942.8

Table B. DD (SR), cross-validated; higher is better. The CA-AVMD row is the reference without intervention.

MethodLlamaMistralQwen
CA-AVMD18.8 (53.5)28.6 (55.6)12.6 (43.7)
BRIDGe (Table 3 in paper)24.7 (49.6)43.4 (51.3)26.0 (39.2)
BRIDGe, TF-IDF gate41.0 (50.2)55.2 (51.1)48.1 (40.5)
Router, TF-IDF64.2 (50.2)62.9 (52.6)63.0 (39.8)

Table C. BRIDGe vs CA-AVMD, paired bootstrap 95% CIs. Δ\DeltaDD is a gain in deferral discrimination; a negative Δ\DeltaSR is a loss in stable retention.

ModelΔ\DeltaDDΔ\DeltaSRΔ\DeltaDD, cross-validated
Llama-3.1-8B-Instruct+9.2 [4.8, 13.6]−4.3 [−5.9, −2.6]+5.9 [1.6, 10.1]
Mistral-7B-Instruct-v0.3+18.0 [13.9, 22.3]−5.0 [−6.9, −3.2]+14.8 [10.7, 18.8]
Qwen2.5-7B-Instruct+15.0 [10.6, 19.5]−3.5 [−5.2, −2.0]+13.4 [8.9, 17.9]

Thank you for the detailed response and the additional experiments. In particular, I appreciate the cross-validated threshold selection and the paired bootstrap CIs (Table C), which show that the DD gain over CA-AVMD remains statistically significant on all three backbones. However, several concerns remain, and some of the new results make them stronger.

  • The held-out relation-type experiment (Table A) addresses part of my question, but the setup is unclear. The no-intervention row (18.8 DD, 53.5 SR) is identical to the full-evaluation numbers in Table 3. This suggests it may not have been re-evaluated on the held-out subset, so it is hard to interpret the comparison with the BRIDGe rows.
  • The response acknowledges that the probe picks up relation-type and entity cues. This affects the framing, not only the wording. The paper’s novelty is gating “on what the model knows, not only on what it sees,” a rule “neither a prompt nor a fine-tuning objective can express.” If the gate largely reads question properties, it is closer to an input-property gate such as CAST. Table B supports this reading. A TF-IDF router on the question text reaches cross-validated DD of 64.2/62.9/63.0 vs. 24.7/43.4/26.0 for BRIDGe at comparable SR, and swapping in a TF-IDF gate improves BRIDGe on every backbone. I acknowledge that on unseen relation types (Table A) the probe gate degrades less than TF-IDF (22.2 vs. 13.9 DD after re-tuning). Still, this advantage is modest, and hard to interpret as the “probe gate” variant is not explained relative to the Table 3 configuration, which underperforms no intervention (15.6 vs. 18.8 DD, 39.8 vs. 53.5 SR). The requested per-relation probe accuracy and semi-volatile split (Table 6) results are also missing. in addition, none of the new tables report CO, BRIDGe’s main stated advantage. This matters especially for the router, which replaces outputs with a deferral string.
  • Recasting the contribution as a “lightweight, interpretable intervention demonstrating coupling” is narrower than the manuscript’s claims. Also, \tau_v and \tau_c were selected on the evaluation questions and revised during rebuttal, so Table 3 should report the cross-validated numbers, which are lower on all backbones. With some several changes, given it is allowed to upload revised pdf during rebuttal, I would like to see the revised manuscript reflect this. Also, Qwen’s \alpha was also set after \alpha=20 was observed to be inert on the evaluation set, this choice was therefore informed by evaluation results as well.
  • The response states that the paper makes no claim of superiority over CAST or SHIFT. However, the paper proposes a new conditional steering method, and to evaluate it as such, readers need to know what it adds over existing methods in the same family. The presence-only gate is an internal ablation of BRIDGe, so it cannot fill this role. This comparison has become more important after the response. Since the volatility gate is acknowledged to rely mainly on relation and entity cues in the input, CAST, which conditions steering on input properties, is the most direct point of comparison. Without it, it is unclear whether the dual-gate design contributes beyond an established approach.
  • On the other points; Qwen still has no multi-seed results, and the Table 2 baselines are still single runs. The concern about absolute performance (VRR 44–60%, SDR up 3–5x) is acknowledged but not mitigated. Table C confirms that the SR loss is significant on all backbones. The presentation issues were not addressed, including abbreviations, a concrete example per cell of the rule, an explanation of the snapshots in the main text, and consistent model names.
AuthorsOpenReview ↗

CAST Comparison

Thank you for the follow-up.

We applied CAST to our CA-AVMD model. It needs two conditions: volatile and no context. We fitted both condition vectors with CAST’s own search over layer and threshold (best F1 on held-out pairs), and used CAST’s refusal vector (our deferral string vs. the model’s own answer) as the steering direction. CAST was fitted on the same questions as BRIDGe and tuned with the same selection rule.

The no-context condition is perfect, but the volatile one fires on 79–85% of stable questions, against 16–19% for BRIDGe. So CAST either has no effect or costs stable answers, and its DD stays below BRIDGe’s (Table D). This holds against BRIDGe’s cross-validated numbers too (24.7 / 43.4 / 26.0, Table B). Pairing CAST’s gate with BRIDGe’s steering doesn’t help either: the model then defers on 70% of stable questions. So the gap comes from the gate, not from how CAST steers.

We agree that BRIDGe belongs to the same family as CAST, and we will frame it that way rather than as “what the model knows”. What it adds is a volatility gate selective enough to be useful; on these models, CAST’s condition is not.

Table D. CAST vs. BRIDGe on CA-AVMD. SR = stable retention, DD = deferral discrimination, CO = context obedience. "Fires on stable" = how often steering triggers on stable questions.

ModelMethodSRDDCOFires on stable
Llama-3.1-8B-InstructNo intervention53.318.562.7–
BRIDGe50.027.062.716.5
CAST52.618.762.785.1
CAST gate + BRIDGe steering11.222.462.785.1
Mistral-7B-Instruct-v0.3No intervention55.628.662.0–
BRIDGe50.745.062.018.8
CAST46.728.862.083.8
CAST gate + BRIDGe steering6.523.262.083.8
Qwen2.5-7B-InstructNo intervention43.912.254.9–
BRIDGe40.527.454.017.5
CAST43.712.554.979.4
CAST gate + BRIDGe steering4.118.654.779.4

CAST strength chosen with BRIDGe’s selection rule; no strength meets it, so the best DD is shown. All rows come from one re-run with the released code; they differ from Table 3 by up to 1.7 DD (BRIDGe) and 0.4 DD (no intervention).

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

This paper introduces BRIDGe, an inference-time activation steering method designed to improve selective abstention when language models encounter outdated factual knowledge. Although factual volatility and context presence are linearly decodable from model activations, conventional prompting and supervised fine-tuning fail to reliably connect volatility to abstention decisions. BRIDGe addresses this issue by using linear probes to detect both signals and conditionally steering hidden states along the context-presence direction. Experiments on three models, primarily using the EvoWiki benchmark, show that it improves deferral discrimination over context-augmented fine-tuning and a presence-only steering baseline.

Critical strengths

The paper addresses an important problem about how to make LLMs defer when their knowledge is outdated, without unnecessarily refusing to answer stable facts. The proposed approach offers a simple inference-time intervention that avoids model fine-tuning, and its dual-gating mechanism is supported by mechanistic analysis. The evaluation also goes beyond overall deferral rates by distinguishing stable-fact retention, stable deferral, and volatile-fact refusal, providing a more nuanced view of the trade-offs.

Critical weaknesses
  1. The paper defines deferral using a set of phrases and their paraphrases, but the implementation of the deferral detector is not sufficiently clear. It is unclear whether deferral is identified through rule-based matching, a separate classifier, or an LLM-based judge. Each approach may introduce different biases or errors. Since SDR, VRR, and DD all depend on this detector, the authors should clarify its implementation and validate its reliability.

  2. The proposed metrics provide a more informative evaluation than the overall deferral rate alone. However, improvements in DD do not necessarily imply better overall model performance, as increased selective abstention comes at the cost of correctly answering stable facts. In addition, the use of contains-match for answer evaluation may not reliably capture semantic correctness or contradictory answers. The results would be more convincing with additional analysis of these trade-offs and the robustness of the evaluation metrics.

Questions for authors
  1. How is deferral detected, and how do the authors validate that the detection method does not bias the reported results?

  2. How should the improvement in the DD metric be interpreted, given the potential trade-off with stable-fact retention and the limitations of contains-match evaluation?

  3. How well does the snapshot-based definition of volatility generalize to real-world facts whose changes may not be captured by two snapshots?

  4. What is the additional inference-time cost of BRIDGe in terms of latency, throughput, and memory usage?

  5. How do the authors establish that the probes capture generalizable volatility signals rather than dataset-specific patterns or correlations?

2 comments

Thank you for the review and for the questions raised.

Q1. Deferral is detected through rule-based matching; no classifier or large language model (LLM) evaluator is used. We use a list of 36 phrases, including the deferral string "I don't have reliable information" and variants such as "I don't know", "I'm not sure", and "I cannot answer". An output is counted as a deferral if it contains at least one of these phrases. The same list is applied to every model and to both stable and volatile questions. This ensures a consistent detection procedure, although it does not rule out missed deferrals or false positives.

Q2. DD measures whether the model defers more often on volatile questions without context than on stable ones (VRR − SDR). The goal is selective deferral, rather than increasing deferral across all questions. A model that always or never defers has DD = 0, so an increase in DD should not be interpreted as an improvement in overall model performance. We therefore report all five metrics together (Section 2.1). BRIDGe improves DD by 9.2 / 18.1 / 15.1 pp over CA-AVMD, while SR decreases by 4.2 / 5.0 / 3.6 pp, respectively (Table 3, Figure 2b). Whether this trade-off is acceptable depends on the relative consequences of outdated answers and unnecessary deferrals.

DD does not rely on contains-match, as it is computed from deferral rates using the detector described in Q1. Contains-match is used for SR and CO, where it remains a limitation because substring matching does not assess semantic correctness or reliably identify contradictory answers (Section 6).

Q3. In our evaluation, volatility is defined by whether a fact's value changes between two snapshots, following EvoWiki's labeling protocol.

A fact whose change is not captured by these two snapshots is therefore not labeled volatile in our data, and we have not tested generalization to such cases. We will clarify this limitation in the paper.

Q4. The inference cost of BRIDGe is the cost of running the SLM. The two directions and the two thresholds of the gate are fitted once, offline, and are only applied at inference. Once per prompt, before the first output token, the gate computes two dot products and, if both conditions hold, adds one vector to one hidden state. It adds no forward pass.

Q5. Table 4 shows that steering along the volatility direction improves DD on Llama and Mistral, whereas a direction fitted on shuffled labels consistently decreases DD. This supports the presence of useful volatility-associated information, although it may partly reflect relation-type and entity cues.

To test generalization, we evaluated BRIDGe on unseen relation types (Table A). The original gate performs below CA-AVMD (DD 15.6 vs. 18.8 pp), but a logistic-probe gate improves DD to 22.2 pp, with only a 1.5 pp reduction in SR. These results suggest that the signal remains useful beyond the training relation types, although its effective use depends on the gate. Generalization to other domains remains untested.

Table A. Llama, on relation types not seen in training (higher is better).

MethodDDSR
No intervention (CA-AVMD)18.853.5
BRIDGe, probe gate22.252.0
BRIDGe (Table 3 in paper)15.639.8
Router, TF-IDF−4.716.9
Router, TF-IDF, threshold re-tuned13.942.8

Thank you for the rebuttal. I believe my initial rating already reflects my assessment of the paper, and I'd like to keep it.

Minor: I suggest correcting the quotation mark issue in Lines 84-85 of the manuscript.

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

The paper introduces the Behavioral Rule Installed via Dual-Gate steering (BRIDGe) method, which can be integrated into small language models to provide an additional mechanism for deciding when to answer questions and when to defer. The authors show that correct deferral is hard to achieve with current base and even specifically fine-tuned models. However, they show that volatility of facts can be identified by linear probing of the models, showing that this signal is present in the residual stream and could be used. They therefore propose employing an explicit gating mechanism that steers the model to defer when volatility is high and context is low. Their method achieves the highest score for deferral discrimination on the EvoWiki dataset.

Critical strengths

The paper proposes a simple and effective mechanism to increase the deferral rate of small language models when they lack context for volatile facts, while still providing answers reliably for stable facts. This mechanism proves more effective than costly fine-tuning of language models and requires only linear probing and model steering.

Critical weaknesses

The motivation for steering along the "context presence" axis is not yet clear to me. The authors mention that the presence of context can be linearly decoded as well (line 200), but they did not provide evidence for this. I would assume that there could be other steering directions possible to make the model defer from answering. It is not clear to me whether this was considered or not.

Questions for authors
  • Q1: The authors mention that "context presence is linearly decodable as well" (line 200), but if I didn't miss anything, there is no explanation or further evidence for it. Why so?
  • Q2: Is pushing c^\hat{c} the only option to bring the model to defer? Might there be a better axis along which the model defers more reliably?
  • Q3: Could you clarify how exactly you identify v^\hat{v} and c^\hat{c} (also which parts of the data are used, etc.) or point me to where it is in the paper if I overlooked it?
  • Q4: Have you tested bigger language models on their performance on this task out of the box? It would be interesting to see whether the deferral problem appears only in small language models or also in larger ones.
  • Q5: What does "BRIDGe applies the hook at prompt_last when the two conditions hold" (line 234) mean? It seems code-related to me.

Suggestion:

  • You could mention related work [1], which uses logical rules to steer models in multiple applications, including question answering.

[1] Helff, Lukas, et al. "Activationreasoning: Logical reasoning in latent activation spaces."

2 comments

We thank the reviewers for their thoughtful questions. Q1, Q3, and Q5 identify information that was missing or insufficiently clear in the submission. We address these points below and will incorporate the corresponding clarifications and results into the revised paper.

Our motivation follows from the training objective of the model we steer. In the CA-AVMD loss (Eq. 2), context is the only difference between the inputs of a prompt trained to defer and a prompt trained to answer. This motivates the hypothesis that moving the hidden state toward the “no context” representation increases deferral, as further explained by the results in Table 4.

Q1. We apologize for omitting this evidence when shortening the section. A logistic probe trained on the hidden state of the last prompt token distinguishes prompts without context from the same questions paired with a relevant passage. The following table reports 5-fold cross-validated accuracy for the CA-AVMD model:

BackboneProbe layerAcc.BRIDGe layerAcc.
Llama-3.1-8B160.997310.995
Mistral-7B161.000310.995
Qwen2.5-7B120.998270.993

The probe uses 300 prompts per class. We will add these results to Appendix A.5 and reference them in the main text at line 200.

Q2. We do not claim that −ĉ is the only option or that it is the best one. We tested four directions without a gate, at three magnitudes, on three backbones (Table 4): −ĉ, +v̂, a random direction, and a direction fitted on shuffled volatility labels. While +v̂ also raises DD, the two control directions do not. Our claim is that −ĉ is sufficient to increase deferral. Moreover, it requires no deferral labels and can be computed for every backbone.

Q3. Both directions v̂ and ĉ are computed on the CA-AVMD model, from the hidden state of the last prompt token at the hook layer (layer 31 on Llama and Mistral, layer 27 on Qwen).

  • ĉ: the weight vector of the logistic probe of context presence (ℓ2-regularized, balanced classes), ℓ2-normalized. The two classes are the volatile evaluation questions without context and the same questions with their passage.
  • v̂: the difference between the mean hidden state of the volatile and stable evaluation questions, in the prompt without context, ℓ2-normalized.
  • τv is the value at which 30% of stable prompts satisfy the volatility condition. τc is one value per backbone. The directions and the thresholds are computed on the evaluation questions.

Q4. We acknowledge this limitation in the current Limitations section. All experiments are conducted on 7–8B models, and we make no claims about generalization to larger models. We will make this scope explicit in the title and abstract.

Q5. "prompt_last" is an identifier from our code and denotes the last token of the prompt. We will replace the current wording with the following: “Both projections are computed once per prompt, using the hidden state h of the last prompt token at layer l, before the first output token is generated. If both conditions hold, −αĉ is added to h.”

Finally, we thank the reviewer for the reference. We will cite ActivationReasoning (Helff et al., ICLR 2026) in the discussion of conditional steering in Section 7.

I thank the authors for their detailed response and clarifications, which I find sensible and helpful.

Considering also the other reviews and the authors' responses, I think some important questions remain open, particularly regarding the generalizability of volatility detection beyond the evaluated setting. Moreover, the additional comparison with text-based routing presents a somewhat mixed picture and raises questions about the relative advantages of the proposed approach.

Nevertheless, I still find the general idea and framework compelling, especially the use of multiple activation-level signals to implement conditional behavior at inference time. I therefore intend to maintain my positive score.

I would, however, appreciate an updated PDF that incorporates the additional results and clarifications from the rebuttal, particularly a more careful positioning of the contributions and corresponding adjustments to the claims, which would help me argue in its favor during the final discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.