acceptodds
Under review as a conference paper at ICLR 2027

BaRe-Mem: Bayesian Reliability Memory for Robust and Adaptive Agent Consultation

Abstract

In multi-agent systems, reliable consultation is challenging because advisor capabilities vary across tasks, and misleading information can make consultation worse than autonomous reasoning. We introduce **BaRe-Mem**, an online Bayesian reliability memory for multi-agent consultation. It estimates advisor reliability based on the central model's internal belief representations and updates these estimates from historical interactions. These estimates modulate the influence of advisor responses and guide the choice between consultation and autonomous reasoning. Across nine benchmarks and six central models, BaRe-Mem is more robust to misleading advisor information than debate and majority voting. On the more challenging tasks, it remains above autonomous reasoning across all tested misleading levels. Moreover, we extend the BaRe-Mem mechanism to worker allocation in agent teams. On the MuSiQue benchmark, BaRe-Mem improves task completion over routing by historical success counts and identifies capable workers earlier. ***The code is in the Supplementary Material for reproduction***.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multi-Agent Systems, Reliability Modeling, Adaptive Consultation, Agent Teams

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_9WoaOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces BaRe-Mem, an online Bayesian reliability memory mechanism for multi-agent consultation that estimates advisor reliability from the central model's internal beliefs and verified interactions. It addresses two core challenges: how much to weight advisor responses, and when to consult versus reason autonomously. Validated across nine benchmarks, BaRe-Mem shows strong robustness against misleading information and significantly improves task completion efficiency in realistic agent team routing scenarios.

Critical strengths
  1. BaRe-Mem's use of Bayesian linear regression with Kalman gain updates provides mathematically justified uncertainty quantification, adaptive weighting, and exact order-independent online updates. Appendix A.7 connects mathematical properties to behavioral predictions, providing an explanation why stronger autonomous models require more reliable evidence before consultation becomes beneficial.

  2. The experiments are convincing across both capability regimes, showing BaRe-Mem successfully maintains performance by adaptively shifting to autonomous reasoning (e.g., dropping consultation to 21% under 100% misleading info in challenging regimes). Beyond synthetic corruption, the method shows tangible real-world utility in the MuSiQue agent routing application, improving task completion by 13-34% with fewer worker calls. The results support the method's applicability beyond response-level consultation.

Critical weaknesses
  1. The paper lacks critical bounds on when the method fails, such as characterizing performance degradation below the 1% sparse verification mark. It assumes IID corruption and ignores adversarial scenarios such as structured misleading information explicitly targeting the central model's uncertainty, or "sleeper" advisors that build false reliability on easy tasks before failing critical ones. There is also no analysis of how quickly the system detects newly unreliable advisors or recovers when they improve, making it hard for practitioners to assess deployment risks.

  2. Several crucial algorithmic decisions are presented without principled justification or targeted ablations to isolate their contributions. The paper does not explain why frozen pre-consultation representations are optimal compared to post-consultation states, why the attention modulation uses arbitrary logarithmic scaling and max-normalization rather than standard softmax, whether the strict linear assumption between representation and correctness holds in practice, or how sensitive the final system is to miscalibration of the decision threshold parameters (ρ,δ\rho, \delta).

Questions for authors
  1. What is the minimum verification rate (e.g., <1%) needed for stable performance, and how quickly does the method degrade below this critical threshold across different capability regimes?

  2. How does BaRe-Mem perform under adversarial corruption patterns, such as strategically providing misleading information when the central model is uncertain, or poisoning attacks where advisors build false reliability on easy questions?

  3. What is the theoretical or empirical rationale for using pre-consultation frozen representations rather than post-consultation states (which reflect advisor influence) for reliability estimation?

  4. Can you provide ablations justifying the reliability-guided attention design, specifically comparing the logarithmic scaling and max-normalization against alternatives like linear scaling or standard softmax?

  5. Given the strict linear assumption for Bayesian updates, have you explored the performance vs. computation trade-off of using non-linear models (e.g., kernel methods or neural predictors) to capture more complex reliability patterns?

7 comments
AuthorsOpenReview ↗

Response to Weakness 1, Questions 1 and Question 2: (1/2)

We thank the reviewer for asking where the boundaries of our method lie. Evaluating these settings directly establishes three findings: BaRe-Mem degrades smoothly down to 0.1% verification without a critical threshold, stays above autonomous reasoning under uncertainty-targeted, sleeper, and majority attacks, and detects a newly unreliable advisor within 103 to 313 verified questions.


1. Sparse Verification Limits Below 1% Across Capability Regimes

W1 & Q1: Minimum Verification Rate and Degradation Profile. Characterizing performance degradation below the 1% sparse verification mark across different capability regimes.

We extend Figure 5 down to 0%, 0.1%, 0.25%, and 0.5% verification across both capability regimes and three misleading information ratios in Tables R1 and R2. Both the reliability memory and the consultation gate are masked under the same sparse verification budget, evaluated on every 12th question, corresponding to 1,451 questions in the capability-challenging regime and 360 questions in the capability-supported regime.

Table R1: BaRe-Mem accuracy (%) under sparse verification in the capability-challenging regime.

FeedbackQwen3-14B 0%50%100%Phi-4 0%50%100%
0%70.160.850.667.158.248.4
0.1%72.166.465.467.763.360.8
0.25%72.464.561.168.962.957.4
0.5%73.266.963.765.762.958.4
1%73.667.863.466.263.057.5
5%75.068.864.269.964.161.8
10%74.969.564.671.064.862.6
100%75.971.768.171.967.864.9
No consultation68.268.268.264.864.864.8
  • No critical threshold and rapid initial adaptation via Kalman gain: By Equation 5 and Appendix A.2, our Bayesian update acts as a Kalman filter whose gain gt,kg_{t,k} scales with the posterior variance vt,k=xt,k⊤Λ−1xt,kv_{t,k} = x_{t,k}^\top \Lambda^{-1} x_{t,k}; when uncertainty is high in the early stage or under sparse verification, the effective step size γvt,k1+γvt,k\frac{\gamma v_{t,k}}{1 + \gamma v_{t,k}} is automatically large, enabling rapid adaptation from very few verified questions. Consequently, while 0% verification reduces to unadapted consultation as a lower limit, every verification rate from 0.1% to 100% improves accuracy over 0% verification by +3.7% to +10.9% at 50% misleading information and by +9.0% to +17.5% at 100% misleading information, and the verification rate needed to match No consultation grows smoothly with the misleading information ratio (0% under honest advisors, 5% to 10% at 50% misleading information, and 100% at 100% misleading information).

Table R2: BaRe-Mem accuracy (%) under sparse verification in the capability-supported regime.

FeedbackQwen3-14B 0%50%100%Phi-4 0%50%100%
0%77.576.173.673.672.868.9
0.1%77.576.173.972.571.167.5
0.25%77.576.475.074.770.870.0
0.5%78.976.475.874.472.569.4
1%79.776.476.172.271.970.3
5%78.376.475.675.073.972.5
10%76.975.677.275.075.070.0
100%78.676.475.677.574.471.9
No consultation75.375.375.370.870.870.8
  • Insensitivity in the capability-supported regime: Because strong central model competence and broadly available useful advisor information make consultation beneficial even without memory, accuracy varies little across verification rates and stays above No consultation at 75.3% on Qwen3-14B and 70.8% on Phi-4 across nearly all settings down.
AuthorsOpenReview ↗

Response to Weakness 1, Questions 1 and Question 2: (2/2)

2. Robustness Under Adversarial Corruption Patterns

W1 & Q2: Adversarial Corruption and Sleeper Advisors. Evaluating BaRe-Mem when advisors provide misleading information targeting the central model’s uncertainty, act as sleeper advisors that build false reliability before failing, or form misleading majorities.

In Table R3, we evaluate four structured patterns on the first 6,000 questions of the capability-challenging regime with Qwen3-14B: uncertainty-targeted, where all advisors mislead where the central model is least sure of its own answer; oracle-targeted, where all advisors mislead on the exact questions where the central model is wrong on its own; sleeper advisor, where the highest-rated advisor is honest for the first half of the stream and then misleads where the central model is wrong; and misleading majority, where 4 or 5 of the 6 advisors mislead on every question.

Table R3: Accuracy (%) under adversarial corruption patterns on Qwen3-14B, where Harm % denotes questions where BaRe-Mem consults and loses a correct autonomous answer.

Corruption patternNo consultationQuestion + PeersAdvisors + memoryBaRe-MemConsult %Harm %
Random corruption, 100%78.954.956.079.67.80.2
Oracle-targeted78.974.576.977.578.62.2
Uncertainty-targeted78.968.370.680.364.80.5
Sleeper advisor78.977.379.980.291.93.8
4/6 advisors misleading78.965.878.080.979.02.1
5/6 advisors misleading78.960.474.880.665.71.8
  • Gains under uncertainty-targeted, sleeper, and majority attacks: BaRe-Mem stays strictly above No consultation at 78.9% under the uncertainty-targeted attack by +1.4% at 80.3%, the sleeper attack by +1.3% at 80.2%, and the 4/6 and 5/6 misleading majorities by +2.0% and +1.7% at 80.9% and 80.6%, whereas Question + Peers drops to 60.4%–77.3%, losing 1.6% to 18.5% relative to No consultation.
  • Boundary under oracle-targeted corruption: The oracle-targeted pattern marks the structural limit: because advisors mislead on every question where consultation could rescue an autonomous error, consultation has almost nothing to rescue by construction. Even here, BaRe-Mem limits harm to 2.2% and stays within 1.4% of No consultation at 77.5%, compared with a 4.4% drop for Question + Peers at 74.5%.

3. Detection and Recovery Under Mid-Stream Advisor Shifts

W1: Detection and Recovery Dynamics. Measuring how quickly BaRe-Mem detects a newly unreliable advisor or recovers when an advisor improves.

In Table R4, the most trusted advisor switches behavior mid-stream on the same 6,000-question stream, with forgetting factor ω∈{1,0.999,0.99}\omega \in \{1, 0.999, 0.99\}. Detection delay counts verified questions until its reliability estimate crosses the median of the other advisors; estimate after is the mean over the next 1,000 questions; and accuracy after change is measured over the first 500 questions after the switch.

Table R4: Mid-stream advisor shift dynamics and accuracy (%) on Qwen3-14B.

Advisor changeForgetting ω\omegaDetection delayEstimate beforeEstimate afterBaRe-Mem after changeNo consultationStable-period accuracy
Honest → misleading13130.580.3079.479.681.3
Honest → misleading0.9992770.560.2880.679.680.5
Honest → misleading0.991030.520.3479.279.680.0
Misleading → honest1> 3,0000.220.4882.279.680.1
Misleading → honest0.999> 3,0000.270.5381.679.679.1
Misleading → honest0.99> 3,0000.340.5281.279.678.4
  • Fast detection and conservative trust restoration: A newly misleading advisor is detected within 313 verified questions at ω=1\omega=1, or within 103 verified questions at ω=0.99\omega=0.99 with 1.3% lower stable-period accuracy before the switch at 80.0% versus 81.3%. When a misleading advisor becomes honest, its estimate rises from 0.22 to 0.48 within 1,000 questions without crossing the peer median before the stream ends, showing that trust is withdrawn faster than it is restored. In both directions, accuracy after the change stays within 0.4% of No consultation or exceeds it by up to +2.6%.
AuthorsOpenReview ↗

Weakness 2 + Questions 3, 4, 5: (1/2)

We thank the reviewer for examining these algorithmic choices. Each design choice follows directly from the requirements of online consultation: pre-consultation frozen representations already encode all advisor answers while preserving causal ordering and an exact posterior; logarithmic max-normalization performs exact multiplicative reweighting inside softmax and outperforms alternative attention rules; the linear Bayesian model in representation space surpasses kernel and nearest-neighbor predictors with sub-2 ms updates; and the consultation gate tolerates perturbations to (ρ^,δ^)(\hat{\rho}, \hat{\delta}) of up to ±0.10\pm 0.10 within 1.5%.


1. Rationale for Frozen Pre-Consultation Representations

W2 & Q3: Pre-Consultation Frozen Representations vs. Post-Consultation States. Explaining the theoretical and empirical rationale for estimating reliability from frozen pre-consultation representations rather than post-consultation states.

Our pre-consultation belief representations already reflect advisor influence: as described in Section 3.1, the frozen central model encodes each candidate response in the context of the question after reading all advisor answers. Extracting these representations from the frozen model before generating the consulted answer satisfies three requirements:

  • Causal ordering: Reliability estimates must exist before consultation generation so they can modulate advisor attention in Equation 7 and gate consultation in Equation 12. By contrast, post-consultation states exist only after generation and only on questions where the gate chooses to consult.
  • Unconfounded targets: Post-consultation hidden states are already altered by the memory’s own attention bias βt,k\beta_{t,k}, confounding reliability estimation with the current steering weights.
  • Exact, order-invariant posterior: Keeping the feature extractor frozen maintains a fixed feature space across the stream, ensuring that the Gaussian posterior in Appendix A.4 remains exact and invariant to question arrival order.

2. Justification and Ablations of Reliability-Guided Attention

W2 & Q4: Logarithmic Scaling and Max-Normalization. Justifying the logarithmic scaling and max-normalization in reliability-guided attention and comparing against linear scaling and standard softmax.

Adding βt,k=γlog⁡(pt,k/max⁡k′pt,k′)\beta_{t,k} = \gamma \log(p_{t,k} / \max_{k'} p_{t,k'}) to pre-softmax attention logits in Equation 7 is equivalent to multiplicative power reweighting inside standard attention softmax:

softmax⁡(s+γlog⁡c)=Normalize⁡(softmax⁡(s)⊙cγ),ck=pkmax⁡jpj.\operatorname{softmax}(s + \gamma \log c) = \operatorname{Normalize}(\operatorname{softmax}(s) \odot c^\gamma), \qquad c_k = \frac{p_k}{\max_j p_j}.

Here, max-normalization sets log⁡cmax⁡=0\log c_{\max} = 0 so the highest-rated advisor retains its natural attention weight, and reduces to zero modification across all tokens whenever advisor estimates are equal, such as at initialization.

Table R5: Comparison of reliability-guided attention designs in the capability-challenging regime, reported as Advisors + memory / BaRe-Mem accuracy (%).

Attention designQwen3-14B 0%100%Phi-4 0%100%
Log-ratio + max-normalization, γ=3\gamma=3, ours75.7 / 75.955.1 / 68.172.3 / 71.952.7 / 64.9
Without max-normalization73.9 / 74.254.9 / 67.567.6 / 67.554.2 / 63.5
Linear scaling73.3 / 73.750.9 / 67.170.8 / 70.748.9 / 64.2
Softmax over advisors75.2 / 75.356.6 / 68.169.3 / 69.754.4 / 64.2
Log-ratio, γ=1\gamma=174.5 / 74.852.4 / 67.070.4 / 70.050.3 / 63.7
Log-ratio, γ=2\gamma=275.3 / 75.754.2 / 67.871.3 / 70.951.8 / 63.8
Log-ratio, γ=5\gamma=575.3 / 75.956.0 / 68.873.2 / 73.154.6 / 65.6
  • Consistent gains over alternative attention rules: At γ=3\gamma=3, logarithmic scaling with max-normalization achieves the highest BaRe-Mem accuracy in every column, outperforming removing max-normalization by up to +4.4% on Phi-4 at 0% misleading information, linear scaling by up to +2.2% on BaRe-Mem and +4.2% on Advisors + memory, and softmax over advisors by up to +2.2% on Phi-4 at 0% misleading information. Accuracy also increases smoothly from γ=1\gamma=1 to γ=5\gamma=5, where γ=5\gamma=5 exceeds our fixed default γ=3\gamma=3 by up to 1.2%, confirming that our results do not rely on hyperparameter tuning. Appendix D with Table 15 of the revised paper reports the sensitivity to γ\gamma and λ\lambda in blue.
Reviewer_jCBQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper's contribution is an inference-time coordination method with external states and online updates. It uses online Bayesian linear regression over frozen-model representations to estimate the reliability of advisors and the central model. These estimates guide attention to advisor responses and selection between consultation and autonomous reasoning. The paper evaluates robustness to misleading advice.

Critical strengths
  1. The method connects reliability estimation, response weighting, and consultation selection. Its closed-form updates make the external memory straightforward to maintain, although these updates are standard Bayesian regression machinery.
  2. The paper provides broad evidence for robustness. On the capability-challenging tasks, BaRe-Mem remains above autonomous reasoning across all tested misleading-advice levels. This supports the practical value of combining historical reliability estimates with an autonomous fallback.
Critical weaknesses
  1. The added value of BaRe-Mem is still unclear. Credibility-aware attention and selective collaboration have been explored before (CrAM; SELENE), as has online model selection (Online Multi-LLM Selection via Contextual Bandits). However, the paper does not compare against strong, independent online-learning methods that can also use historical feedback. The success-count baseline in the routing experiment learns from feedback, but ignores the current question. The always-consult ablation shows that falling back to the model’s own answer helps. It does not show how much the Bayesian design adds over a simpler predictor. This makes it hard to tell whether the gains come from BaRe-Mem’s specific design or mainly from using past feedback and allowing autonomous fallback.

  2. The compute and feedback budgets are not clearly accounted for. The main consultation experiments compare accuracy without showing that the methods use comparable model-call or compute budgets. BaRe-Mem requires extra model passes to extract features, and its released evaluation pipeline generates both autonomous and consultation answers. These costs should be included when comparing it with other methods, so we can tell whether the gains remain at similar budgets. Figure 5 should also clarify whether its feedback budget includes the consultation gate’s updates, which are not masked in the released pipeline.

Questions for authors
  1. Have you considered a broader set of baselines in the corresponding consultation and routing tasks? These include contextual online learners such as LinUCB, contextual Thompson sampling, and online logistic regression; selective collaboration methods (SELENE; Free-MAD); and learned routers (Online Multi-LLM Selection via Contextual Bandits; RouteLLM; MasRouter). Which have been evaluated under comparable feedback, training-data, and compute budgets?
  2. What exact procedure produces ψq\psi_q and ψc\psi_c? Please specify the extraction model, prompts, layers, token selection or pooling, normalization, and dimensions before and after projection.
  3. Does the “question” feature use only the question, or also the candidate answers? The released consultation code appears to use the latter. Has a question-only version been evaluated?
  4. Have you tested different feature dimensions, selected layers, pooling methods, and projection methods? How sensitive are the results to these choices?
  5. Have you separately ablated the feature blocks and posterior uncertainty? Beyond the always-consult ablation, have you tested a gate-only variant without attention reweighting, using the same feedback?
  6. What is the practical significance of the injected-noise setting? For example, does robustness to injected noise translate to robustness in a larger advisor pool containing weaker models? Has this or any other pratical cases been tested using their original outputs without artificial corruption?
10 comments
AuthorsOpenReview ↗

Response to Weakness 1 and Question 1: (1/3)

We have evaluated across both consultation and routing compared with these baselines. The results confirm that past feedback or autonomous fallback alone cannot reproduce our gains: methods with identical feedback and fallback still fall below autonomous reasoning unless reliability is estimated at the individual answer level to jointly guide attention and consultation selection.

1. Added Value of BaRe-Mem and Reliability Estimation Quality

W1: Source of Gains and Value of the Bayesian Design. Clarifying whether gains stem from BaRe-Mem’s specific design or merely from past feedback and autonomous fallback, and isolating what the Bayesian formulation adds over simpler predictors.

Prior methods address only isolated pieces of the consultation problem: CrAM [1] reweights attention using external credibility scores without online feedback or abstention, SELENE [2] gates debate by current peer agreement without persistent memory, and contextual bandits [3] select models from the query before any candidate answer exists. BaRe-Mem uniquely estimates the reliability of each candidate answer, including the central model’s own response, within the central model’s belief space from verified history, using one unified estimate to steer attention in Equation 7 and decide whether to consult in Equation 12.

Table R1 directly isolates reliability estimation AUC and Top-1 highest-rated answer selection accuracy among disagreeing advisors on both models under identical verified feedback.

Table R1: Reliability estimation AUC and Top-1 answer selection accuracy on disagreeing advisors under identical verified feedback on the challenging suite at 50% misleading advice.

Reliability estimatorQwen3-14B AUCPhi-4 AUCQwen3-14B Top-1Phi-4 Top-1
Random selection0.5000.50037.1%37.1%
Success counts per advisor0.5490.54943.3%43.3%
Success counts per advisor and task type0.6820.68250.5%50.5%
True advisor and task accuracy, post hoc0.6840.68450.6%50.6%
BaRe-Mem (logistic regression estimator)0.9200.90186.9%83.1%
BaRe-Mem (delta rule estimator)0.9220.89786.7%81.5%
BaRe-Mem (Bayesian estimator, default)0.9530.93590.3%87.3%
  • Historical feedback without belief features hits a low ceiling: The success-count baseline conditioned on advisor identity and task type reaches only 0.682 AUC online and 50.5% Top-1 selection accuracy, bounded by a hard ceiling of 0.684 AUC and 50.6% Top-1 accuracy even with post hoc true accuracies. Incorporating BaRe-Mem’s belief features lifts AUC above 0.89 and Top-1 accuracy above 81% on both models.
  • The Bayesian formulation outperforms simpler predictors on both models: Given the exact same belief features and verified feedback, the Bayesian estimator achieves the strongest discrimination on both Qwen3-14B at 0.953 AUC and Phi-4 at 0.935 AUC, beating online logistic regression by +0.033 and +0.034 AUC and the online delta rule by +0.031 and +0.038 AUC. When selecting the highest-rated candidate answer directly, the Bayesian estimator reaches 90.3% on Qwen3-14B and 87.3% on Phi-4, surpassing online logistic regression by +3.4 pp and +4.2 pp and the delta rule by +3.6 pp and +5.8 pp while providing an exact, order invariant update in Appendix A.4 without learning rate tuning.

In the revised paper, Appendix B.8 with Table 14 reports the estimator comparison of Table R1; all revisions are marked in blue.

References

[1] Deng, Boyi, et al. "Cram: Credibility-aware attention modification in llms for combating misinformation in rag." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 39. No. 22. 2025.

[2] Verma, Akshay, et al. "SELENE: Selective and evidence-weighted LLM debating for efficient and reliable reasoning." Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 5: Industry Track). 2026.

[3] Poon, Manhin, et al. "Online multi-llm selection via contextual bandits under unstructured context evolution." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 40. No. 29. 2026.

AuthorsOpenReview ↗

Response to Weakness 1 and Question 1: (2/3)

2. Broader Baselines in the Main Consultation Setting

Q1: Contextual Online Learners, Selective Collaboration, and Learned Routers. Evaluating these baselines under comparable feedback, training-data, and compute budgets.

In Table R2, SELENE [2] and Free-MAD [4] use their original feedback-free debate protocols. All online-learning baselines start under identical verified feedback with no extra training-data: [3, 5, 6] select among the seven candidate answers including the central model’s own answer; online logistic regression replaces only our Bayesian estimator inside our steering and gate; and LinUCB [7] and Thompson sampling [8] keep BaRe-Mem’s posterior of Equations 4–5 and change only the read-out of Equation 6:

pt,kUCB=Φ(μt,k+vt,k1+vt,k),pt,kTS=Φ(μt,k+vt,kϵt,k),ϵt,k∼N(0,1),p_{t,k}^{\mathrm{UCB}}=\Phi\Big(\frac{\mu_{t,k}+\sqrt{v_{t,k}}}{\sqrt{1+v_{t,k}}}\Big), \quad p_{t,k}^{\mathrm{TS}}=\Phi\big(\mu_{t,k}+\sqrt{v_{t,k}}\epsilon_{t,k}\big), \quad \epsilon_{t,k}\sim\mathcal N(0,1),

with μt,k\mu_{t,k} and vt,kv_{t,k} of Equation 6.

Online logistic regression uses a diagonal Laplace posterior:

q←q+y^t,k(1−y^t,k)xt,k2,m~←m~−(y^t,k−yt,k)xt,kq,pt,kLogReg=σ(xt,k⊤m~1+πv~t,k/8),q \leftarrow q + \hat y_{t,k}(1-\hat y_{t,k}) x_{t,k}^2, \quad \tilde m \leftarrow \tilde m - \frac{(\hat y_{t,k}-y_{t,k}) x_{t,k}}{q}, \quad p_{t,k}^{\mathrm{LogReg}}=\sigma\Big(\frac{x_{t,k}^\top \tilde m}{\sqrt{1+\pi \tilde v_{t,k}/8}}\Big),

with y^t,k=σ(xt,k⊤m~)\hat y_{t,k}=\sigma(x_{t,k}^\top \tilde m) and v~t,k=∑jxt,k,j2/qj\tilde v_{t,k}=\sum_j x_{t,k,j}^2/q_j.

Table R2: Consultation accuracy (%) on the challenging suite across misleading ratios.

ModelMethod0%25%50%75%100%
Qwen3-14BNo consultation66.666.666.666.666.6
Question + Peers69.865.660.455.250.6
SELENE [2]63.452.635.818.69.1
Free-MAD [4]67.356.133.610.02.8
Online Multi-LLM Selection [3]69.965.765.465.565.2
RouteLLM [5]72.368.065.865.365.0
MasRouter [6]60.154.448.642.738.1
BaRe-Mem (logistic regression estimator)74.672.071.470.268.4
BaRe-Mem (UCB read-out) [7]75.272.170.869.567.8
BaRe-Mem (Thompson read-out) [8]75.672.471.369.467.9
BaRe-Mem (Bayesian estimator, default)76.072.871.469.968.6
Phi-4No consultation63.463.463.463.463.4
Question + Peers65.862.457.952.548.3
SELENE [2]63.454.237.721.614.7
Free-MAD [4]65.353.631.08.62.3
Online Multi-LLM Selection [3]69.964.865.464.664.2
RouteLLM [5]72.869.065.362.762.1
MasRouter [6]58.252.446.741.036.3
BaRe-Mem (logistic regression estimator)70.669.768.867.766.3
BaRe-Mem (UCB read-out) [7]70.767.966.565.564.5
BaRe-Mem (Thompson read-out) [8]71.168.466.565.364.6
BaRe-Mem (Bayesian estimator, default)71.268.566.865.365.0
  • Feedback and fallback outside BaRe-Mem are insufficient: [3, 5, 6] use identical feedback and fallback but score sources from the question without inspecting answers: on Qwen3-14B all three fall below No consultation from 50% misleading onward, as do [5, 6] on Phi-4 at 75–100%.
  • Fallback without steering, or debate without memory, fails: Gating unweighted consultation stays at or below No consultation (65.1% / 66.6% in Q5), whereas always-consult with reliability attention adds 7.7–19.9% over Question + Peers (Appendix B.5). Without verified memory, SELENE and Free-MAD drop to 9.1% and 2.8% at 100% misleading.
  • Online learners inside BaRe-Mem’s framework: Reusing our features, steering, and gate keeps [7, 8] and logistic regression above No consultation at all ratios. Because verified feedback labels all candidates regardless of the gate, no exploration is needed; the UCB bonus and Thompson noise only perturb Equation 6 and stay at or below it in 19 of 20 settings. While online gate fitting (Equations 10–11, Appendix A.8) masks score errors to let logistic regression edge ahead on Phi-4 (66.3% vs. 65.0%), the Bayesian estimator leads on Qwen3-14B and is more accurate on both models (+0.033 / +0.034 AUC and +3.4 / +4.2% Top-1 in Table R1) with no learning-rate tuning.

In the revised paper, Appendix B.7 (Table 13) reports the gate-only ablation in blue.

AuthorsOpenReview ↗

Response to Weakness 1 and Question 1: (3/3)

3. Agent Team Routing Comparisons on MuSiQue

The agent team experiment in Section 4.4 serves as an real-world application study testing whether our reliability memory transfers directly to worker allocation in agent team pipeline, rather than the primary task BaRe-Mem is designed for. In Table R3, reliability must be predicted from the sub-task alone before any worker report is generated.

Table R3: Agent team routing on MuSiQue under dataset verifier checking as sub-task and task accuracy in percentage points and worker calls per sub-task.

MethodQwen3-14B Sub-taskTaskWorker CallsPhi-4 Sub-taskTaskWorker Calls
Random routing70.843.01.4670.542.81.46
Success counts75.750.41.3575.450.51.37
LinUCB [7]78.454.01.3078.554.21.29
Linear Thompson [8]77.652.91.3377.552.91.33
Online logistic regression77.352.41.3176.651.61.32
Contextual Bandits, Poon et al. [3]79.054.91.2878.554.61.29
RouteLLM [5]76.651.51.3275.049.81.35
MasRouter [6]71.744.11.4472.446.31.43
BaRe-Mem78.254.31.3078.354.21.29
  • Strong transfer to a downstream routing application alongside core consultation capabilities: Although agent team routing is only a downstream validation scenario for BaRe-Mem and predicting reliability before report generation is the exact setting that dedicated question-level routers [3] are built for, BaRe-Mem, reading only the lead model’s hidden states without external encoders, still comes within 0.8 % of Contextual Bandits [3] on Qwen3-14B and 0.2 % on Phi-4, while outperforming online logistic regression by +0.9 % and +1.7 %, RouteLLM [5] by +1.6 % and +3.3 %, MasRouter [6] by +6.5 % and +5.9 %, and historical success counts by +2.5 % and +2.9 %. Crucially, in our primary consultation setting where candidate responses exist, the same BaRe-Mem memory rates individual answers, steers attention, and decides whether to consult in Table R2, where question-level routers fail.

References

[4] Cui, Yu, et al. "Free-mad: Consensus-free multi-agent debate." Findings of the Association for Computational Linguistics: ACL 2026. 2026.

[5] Ong, Isaac, et al. "Routellm: Learning to route llms from preference data." International Conference on Learning Representations. 2025.

[6] Yue, Yanwei, et al. "Masrouter: Learning to route llms for multi-agent systems." Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics. 2025.

[7] Li, Lihong, et al. "A contextual-bandit approach to personalized news article recommendation." Proceedings of the 19th international conference on World wide web. 2010.

[8] Agrawal, et al. "Thompson sampling for contextual bandits with linear payoffs." International conference on machine learning. PMLR, 2013

Reviewer_A2x8OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes BaRe-Mem, an online Bayesian reliability memory for multi-agent consultation. Hidden-state belief representations of the question and of each candidate response, taken from a frozen central model, feed a Bayesian linear regression, Eqs. 1-6, that estimates per question the probability that each advisor, and the model's own answer, is correct. These estimates serve two purposes. First, they steer attention toward more reliable advisors by adding reliability-dependent biases to pre-softmax attention logits without adding trainable parameters to the attention intervention, as in Eq. 7. Second, they decide whether to consult at all by comparing an estimated consultation ability against the estimated autonomous ability kappa, as in Eqs. 8-13. The evaluation covers 9 benchmarks, 6 central models, and 6 advisors. Tables 4-9 show accuracy above autonomous reasoning at all tested misleading ratios in the capability-challenging regime; this is not universal in the supported regime, where Table 7 reports 69.3% versus a 70.6% autonomous baseline for Qwen2.5-7B at 100% misleading advice. Section 4.4 extends the mechanism to worker routing on MuSiQue.

Critical strengths
  1. The method combines an interpretable online reliability model with a useful consult-or-reason-alone decision. The Gaussian posterior and rank-one updates are derived explicitly in Appendix A, and the attention intervention adds no trainable parameters to the frozen central model. Reliability and consultation parameters are nevertheless learned through online Bayesian updates. The consultation ratio remains high in the supported regime but falls from 90% to 21% for Qwen3-14B in the challenging regime as misleading information increases.

  2. The empirical coverage is broad: nine benchmarks, six central models, six advisors, and an additional MuSiQue routing study with 2,417 tasks and 6,404 sub-tasks. The main and appendix tables report consistent corresponding results. Section B.4 also acknowledges a gap between predicted reliability and empirical correctness under heavily misleading advice, which is useful for assessing the method's limits.

Critical weaknesses
  1. As submitted, the paper does not report enough run-level information to assess statistical stability. Tables 1 and 4-9 give point estimates without seed counts, error bars, or paired tests. The reported 1.5-2.8-point advantage at 100% misleading advice is a challenging-suite aggregate, while the 1.7-3.5-point sub-task advantage is measured on MuSiQue under dataset-verifier checking; both need uncertainty estimates appropriate to their evaluation units and online feedback dependencies. For the two main models, consultation falls to 18-21% only in the challenging regime, while remaining at 83-85% in the supported regime. Falling back to autonomous answers can mitigate harmful advice, but it does not guarantee accuracy above the autonomous baseline: the selector may still choose consultation on the wrong examples. A decomposition of actual rescue and harm on the selected examples would clarify the tail gains.

  2. The contribution relative to prior reliability-aware methods and simple feedback-matched alternatives remains insufficiently isolated in the main consultation experiment. The cited Sigma-Mem is a relevant online reliability-memory comparator, and CRAM is related through credibility-based attention reweighting, although its task, score source, and head-selection mechanism differ. Neither is included as a baseline. In Section 4.2, debate and majority voting do not use the verified online feedback supplied to BaRe-Mem. Section 4.4 does include a historical-success-count router under matched checking and worker budgets; an analogous simple feedback-matched comparison in the consultation setting would help identify the value of the Bayesian formulation.

Questions for authors
  1. Can you report repeated-run and stream-order uncertainty, or appropriately paired tests, for the challenging-suite tail gains and the MuSiQue sub-task gains? Please state the resampling unit and account for feedback dependence.

  2. Can you compare the main consultation setting with Sigma-Mem or an adapted CRAM baseline, and with a simple router given the same belief features and correctness feedback?

  3. Please specify the layer, token positions, and dimensions used for psi_q and psi_c; confirm that current verified labels are unavailable when these features and predictions are formed; and provide values and sensitivity results for gamma and lambda, plus the concrete misleading-response generation protocol.

6 comments
AuthorsOpenReview ↗

Response to Weakness 1 and Question 1(1/2)

We thank the reviewer for emphasizing uncertainty quantification and a fine-grained breakdown of consultation decisions. Both analyses confirm that BaRe-Mem’s tail gains are statistically robust and driven by competence comparison (A(Tt)≥κtA(T_t) \ge \kappa_t) rather than passive fallback.


1. Stream-Order and Within-Stream Uncertainty

W1 & Q1: Reporting repeated-run and stream-order uncertainty for the challenging-suite tail gains and MuSiQue sub-task gains, stating the resampling unit and accounting for feedback dependence.

Because BaRe-Mem’s Gaussian posterior is order-invariant for any fixed observation set (Appendix A.4), stream order affects only early updates before posterior concentration. We evaluate uncertainty across three levels that preserve sequential feedback dependencies:

  • Across-stream order (unit = complete stream, N=17,403N=17,403): Repeating the process across 3 random stream orders yields a standard deviation of at most 0.26 pp across all misleading ratios as shown in Table R1.
  • Within-stream blocks (unit = contiguous 50-question block): A paired circular block bootstrap [1] preserves local feedback dependencies within a run (intervals remain stable across block sizes B∈{25,50,100}B \in \{25, 50, 100\}). The 95% confidence intervals excluded zero in every setting, even when at 100% misleading advice. Under that condition, the estimated gains were +1.48 (95% CI [1.23, 1.74]) for Qwen3-4B and +2.14 (95% CI [1.87, 2.41]) for Qwen3-8B, and McNemar’s tests [2] confirm strong statistical significance across every setting (p<10−21p < 10^{-21}).
  • MuSiQue routing (unit = multi-hop task, N=2,417N=2,417): We average outcomes across 3 stream orders and apply a paired task bootstrap keeping all sub-tasks of the same task together. As shown in Table R2, all 95% CIs for both sub-task and task gains are above zero.

Table R1: Challenging-suite gains (pp) of BaRe-Mem over No consultation.

Central ModelMisleadingGain (Seed 0)95% Block CI (B=50B=50)McNemar pp3 Order Seeds (Mean ± SD)
Qwen3-14B0%+9.38[8.79, 9.96]3.0×10−2323.0 \times 10^{-232}+9.35 ± 0.07
Qwen3-14B25%+6.14[5.57, 6.66]1.7×10−1151.7 \times 10^{-115}+6.25 ± 0.19
Qwen3-14B50%+4.81[4.30, 5.31]1.4×10−851.4 \times 10^{-85}+4.83 ± 0.13
Qwen3-14B75%+3.25[2.82, 3.68]4.0×10−534.0 \times 10^{-53}+3.31 ± 0.09
Qwen3-14B100%+1.94[1.65, 2.25]1.2×10−421.2 \times 10^{-42}+1.91 ± 0.06
Phi-40%+7.87[7.33, 8.38]6.9×10−2006.9 \times 10^{-200}+7.88 ± 0.04
Phi-425%+5.17[4.65, 5.65]2.8×10−1002.8 \times 10^{-100}+5.31 ± 0.13
Phi-450%+3.40[2.92, 3.87]2.5×10−522.5 \times 10^{-52}+3.64 ± 0.21
Phi-475%+1.93[1.52, 2.33]5.2×10−225.2 \times 10^{-22}+2.19 ± 0.26
Phi-4100%+1.61**[1.33, 1.89]8.2×10−438.2 \times 10^{-43}+1.66 ± 0.12

Table R2: MuSiQue routing gains (pp) under dataset-verifier checking across 3 stream orders.

Lead ModelComparisonSub-task Gain (pp)95% Task CITask Gain (pp)95% Task CI
Qwen3-14Bvs. Success counts+2.52[1.98, 3.06]+3.93[2.87, 5.03]
Qwen3-14Bvs. Random routing+7.45[6.87, 8.07]+11.32[10.19, 12.50]
Phi-4vs. Success counts+2.92[2.26, 3.53]+3.70[2.43, 4.91]
Phi-4vs. Random routing+7.80[7.20, 8.42]+11.38[10.16, 12.56]

References

[1] Politis, Dimitris N., and Joseph P. Romano. A circular block-resampling procedure for stationary data. Purdue University. Department of Statistics, 1991.
[2] McNemar, Quinn. "Note on the sampling error of the difference between correlated proportions or percentages." Psychometrika 12.2 (1947): 153-157.

AuthorsOpenReview ↗

Response to Weakness 1 and Question 1(2/2)

2. Consultation Adaptation and Rescue–Harm Decomposition

W1: Clarifying the 83–85% (supported) vs. 18–21% (challenging) consultation rates, and decomposing selected examples into rescues vs. harms.

Why consultation stays at 83–85% in the supported regime versus 18–21% in the challenging regime.

This contrast demonstrates our core conclusion: BaRe-Mem adapts consultation frequency to the actual degradation risk δ^\hat{\delta} of unreliable advice and the resulting break-even threshold T∗T^* . In the supported regime, strong baseline competence and broadly available useful advisor information in Table 2 make the central model resilient for Qwen3-14B at 100% misleading, even always-consulting Advisors + memory still outperforms No consultation at 74.5% versus 72.9% in Table 6 when keeping δ^\hat{\delta} small and high consultation optimal. In the challenging regime, heterogeneous and task-dependent capabilities in Table 3 amplify misleading interference, dropping Advisors + memory to 54.7% versus 66.6%. This large δ^\hat{\delta} raises T∗T^* and prompts BaRe-Mem to cut consultation to 20.8%.

Rescue and harm decomposition. On consulted questions, we define Rescue as consultation succeeding while autonomous reasoning fails, and Harm as consultation failing while autonomous reasoning succeeds. By construction, Rescue−Harm\text{Rescue} - \text{Harm} equals the accuracy gain over No consultation.

Table R3: Rescue and harm decomposition on consulted examples at 100% misleading advice as percentages of all questions.

ModelSuiteConsult %Rescue %Harm %Gain %
Qwen3-14BChallenging20.82.80.9+1.9
Phi-4Challenging18.22.10.5+1.6
Qwen3-14BSupported84.75.84.5+1.3
Phi-4Supported83.27.35.7+1.6
Qwen2.5-7BSupported77.05.77.0−1.3
  • Rescues outweigh harms by over 3×3\times in the challenging tail: BaRe-Mem identifies genuinely beneficial consultations with 2.8% rescue versus 0.9% harm for Qwen3-14B and 2.1% versus 0.5% for Phi-4. This aligns with Figure 4, where the predicted consultation advantage Δ^t=A(Tt)−κt\hat{\Delta}_t = A(T_t) - \kappa_t increases monotonically with the empirical gain and crosses zero close to the true decision boundary.
  • Broad consistency across settings: By the identity above, adaptive selection beats autonomous reasoning whenever rescues exceed harms, which holds in 58 of 60 evaluated settings across models, suites, and misleading ratios in Tables 4–9, including all 30 settings in the challenging suite. Even in the Qwen2.5-7B supported tail at 69.3% versus 70.6%, BaRe-Mem recovers +2.2 pp over unguided consultation at 67.1% in Table 7.

We have added both analyses to the revised paper: Appendix B.6 reports the uncertainty analysis in Tables 10 and 11, and Appendix B.7 reports the rescue and harm decomposition in Table 12. All revisions are marked in blue, and we are glad for you to check them.

AuthorsOpenReview ↗

Response to Weakness 2 and Question 2: (1/2)

We thank the reviewer for suggesting these comparisons. Running all baselines in the main consultation setting establishes two results: BaRe-Mem outperforms prior reliability methods across every misleading ratio, and under identical belief features and verified feedback its Bayesian posterior estimates reliability more accurately on both models than historical success counts and simpler online predictors.

W2 & Q2: Prior Methods and Feedback-Matched Baselines. Comparing BaRe-Mem in the main consultation setting against Σ-Mem, an adapted CrAM baseline, and simpler baselines or routers given the same belief features and verified correctness feedback.

1. Main Consultation Comparisons

In Table R4, Σ-Mem [1] and CrAM [2] retain their original mechanisms. The three BaRe-Mem rows share the belief features xt,kx_{t,k} in our Equation 1, the verified labels st,k=2yt,k−1s_{t,k}=2y_{t,k}-1, the attention steering of Equation 7 and the consultation gate of Equations 8–12, and only differ how pt,kp_{t,k} is learned from the labels:

Bayesian (default):m←m+gt,k(st,k−xt,k⊤m),pt,k=Φ(xt,k⊤m1+vt,k),\text{Bayesian (default):}\quad m \leftarrow m + g_{t,k}\big(s_{t,k}-x_{t,k}^\top m\big), \qquad p_{t,k}=\Phi\Big(\frac{x_{t,k}^\top m}{\sqrt{1+v_{t,k}}}\Big),

Delta rule:m←m+βxt,kxt,k⊤xt,k(st,k−xt,k⊤m),pt,k=Φ(xt,k⊤m),\text{Delta rule:}\quad m \leftarrow m + \frac{\beta x_{t,k}}{x_{t,k}^\top x_{t,k}}\big(s_{t,k}-x_{t,k}^\top m\big), \qquad p_{t,k}=\Phi\big(x_{t,k}^\top m\big),

Logistic regression:q←q+y^t,k(1−y^t,k)xt,k2,m←m−(y^t,k−yt,k)xt,kq,pt,k=σ(xt,k⊤m1+πvt,k/8),\text{Logistic regression:}\quad q \leftarrow q + \hat y_{t,k}(1-\hat y_{t,k}) x_{t,k}^2, \quad m \leftarrow m - \frac{(\hat y_{t,k}-y_{t,k}) x_{t,k}}{q}, \qquad p_{t,k}=\sigma\Big(\frac{x_{t,k}^\top m}{\sqrt{1+\pi v_{t,k}/8}}\Big),

where gt,kg_{t,k} is the Kalman gain of Equation 5, β=0.05\beta=0.05, y^t,k=σ(xt,k⊤m)\hat y_{t,k}=\sigma(x_{t,k}^\top m), the square and the division by qq act elementwise, and vt,k=∑jxt,k,j2/qjv_{t,k}=\sum_j x_{t,k,j}^2/q_j for logistic regression. The delta rule is thus Equation 5 with the Kalman gain replaced by a fixed step, and logistic regression replaces the exact Gaussian posterior with a diagonal Laplace approximation.

Table R4: Accuracy in percentage points on the challenging suite across misleading advice ratios.

ModelMethod0%25%50%75%100%
Qwen3-14BNo consultation66.666.666.666.666.6
Question + Peers69.865.660.455.250.6
Σ-Mem61.349.734.116.05.5
CrAM72.667.563.457.852.4
BaRe-Mem (logistic regression estimator)74.672.071.470.268.4
BaRe-Mem (delta rule estimator)75.471.269.168.067.0
BaRe-Mem (Bayesian estimator, default)76.072.871.469.968.6
Phi-4No consultation63.463.463.463.463.4
Question + Peers65.862.457.952.548.3
Σ-Mem63.349.231.214.34.7
CrAM67.665.462.958.154.7
BaRe-Mem (logistic regression estimator)70.669.768.867.766.3
BaRe-Mem (delta rule estimator)70.967.565.163.664.3
BaRe-Mem (Bayesian estimator, default)71.268.566.865.365.0
  • Superiority over prior reliability methods: Σ-Mem [1] selects only among advisor answers without an autonomous fallback option, so its accuracy is bounded by the advisor pool and drops to 5.5% on Qwen3-14B and 4.7% on Phi-4 at 100% misleading advice. CrAM [2] scores current answers without verified history and always consults, falling below No consultation once half of the advice is misleading at 63.4% versus 66.6% on Qwen3-14B and 62.9% versus 63.4% on Phi-4.
  • Effectiveness of BaRe-Mem’s steering and gating architecture: Equipping simpler online predictors with BaRe-Mem’s belief features, attention steering, and consultation gate keeps both variants above No consultation across all ratios, confirming the effectiveness of our dual steering and gating design. Within this shared framework, replacing only the Kalman gain with a fixed step lowers accuracy in all ten settings and lowers AUC and Top-1 on both models in Table R5, which isolates the value of the exact Bayesian update; the Bayesian estimator also leads or ties on Qwen3-14B at 0%, 25%, 50%, and 100% misleading advice as well as on Phi-4 at 0% misleading advice.
  • Why binary gating masks score errors on Phi-4 while BaRe-Mem’s estimator remains fundamentally superior: Because the consultation gate fits ρ^\hat{\rho} and δ^\hat{\delta} online to each estimator’s score scale, binary gate thresholding can mask estimator errors whenever the sign of A(Tt)−κtA(T_t) - \kappa_t is preserved in Appendix A.8, resulting in slightly more conservative abstention for logistic regression on Phi-4 under corruption. Fundamentally, however, the Bayesian estimator provides a substantially more accurate and order invariant reliabili

Comments on the paper

1 comment
AuthorsOpenReview ↗

A gentle reminder to read our rebuttal

Dear Reviewers,

Thank you for your time and effort in reviewing our submission. We are extremely sorry for being late in responding to your comments. As you can see, we had to conduct several experiments and provide theoretical justifications, which took quite some time. This also made our responses very lengthy, and we are extremely sorry for that.

With the discussion period drawing to a close, we kindly invite you to review our responses and the revised manuscript. We have conducted all the additional experiments you requested and incorporated the results into the updated paper. Please let us know if you have any remaining questions or concerns.

If our responses have adequately addressed your feedback, we would be very grateful if you would consider raising your score.

--

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.