acceptodds
Under review as a conference paper at ICLR 2027

Agent-Conditioned Routing: Diverse Reasoning Paths Within One Frozen Model

Abstract

Generating several solutions can expose alternative methods, but additional trajectories often paraphrase one dominant approach. We investigate whether the conditional capacity of a frozen sparse mixture-of-experts (MoE) model can provide such diversity from one problem prompt, without user-assigned roles, prescribed methods, fine-tuning, or model copies. Agent-Conditioned Routing (ACR) assigns parallel trajectories persistent, untrained shared/private expert-eligibility masks while preserving the learned, input-dependent token router. On a screened 126-question multi-mode mathematics and physics benchmark, with model judges reading the reported derivations in full, we count the number of problems yielding at least two distinct primary methods among soft-correct solutions in four trajectories. ACR reaches approximately twice as many such problems as the strongest temperature setting under GLM 5.3 (37.3 versus 18.7), and 1.67 times as many under GPT-5.6 Sol (45.00 versus 27.00), at matched output budgets. A blinded four-expert audit on a stratified sample of 30 problems also finds more multi-method solutions with ACR on both backbones, with lower correctness. Matched-eligibility ablations show modest multi-method count advantages for ACR, while token-resampled masks attain higher correctness. An informed sequential prompt elicits more diversity still, using previous solutions and explicit requests for alternative methods. ACR adds about 160 KiB of mask state. In a three-problem microbenchmark capped at 1,024 tokens, its latency and peak VRAM usage remain within 3% of batched temperature sampling.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Diversity, LLMs, Mixture-of-Experts, Expert-Routing

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_YtjjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Agent-Conditioned Routing (ACR), which assigns persistent shared/private expert-eligibility masks to different trajectories during the generation prefix of a frozen mixture-of-experts model. The goal is to increase solution-method diversity without parameter updates or assigned roles. The authors evaluate ACR on a self-constructed benchmark of 126 mathematics and physics problems using model judges, a blinded human audit, and ablation studies. The results show an increase over temperature sampling in the number of problems with at least two distinct primary methods among answer-correct trajectories, although some configurations incur a correctness loss.

Critical strengths

This work propose a simple, lightweight inference-time diversification mechanism supported by complementary evaluations. The method requires no training or additional model copies and supports parallel generation of independent trajectories. Model-based evaluation, a blinded human audit, and routing ablations jointly provide empirical support for its diversification effect.

Critical weaknesses
  1. The novelty and value of the additional routing design are not sufficiently established. Routing interventions for increasing reasoning diversity have already been explored. The main additional design here is the persistent shared/private expert mask. However, the ablations show that simpler alternatives retain most of the gains, while the full construction provides relatively small additional improvements without paired uncertainty estimates. The evidence therefore does not sufficiently establish the distinctive value of this design.

  2. The necessity and independent value of the self-constructed benchmark are insufficiently justified. Evaluating solution-method diversity requires derivation-level annotation, but not necessarily newly authored problems. The paper does not adequately explain why the same evaluation protocol cannot be applied to public benchmarks such as AIME or LiveCodeBench. The main results are also confined to the self-constructed, screened dataset, without systematic validation on standard benchmarks. Consequently, both the benchmark's independent contribution and the method's external validity remain insufficiently established.

Questions for authors
  1. What is the contribution of the additional mask structure? In Table 7, simpler masking schemes retain most of the gains. Are the additional improvements from persistent shared/private masks stable? How should their diversity–correctness tradeoff relative to token-resampled masks be interpreted?

  2. Why is a self-constructed benchmark necessary? What specific limitations of existing mathematical benchmarks prevent the same protocol from being applied to AIME or LiveCodeBench? Are there results on standard benchmarks that support the method's generalizability?

  3. How do you ensure that the evaluation captures diversity of valid solution methods? Correctness is primarily determined by whether the correct answer is reached. Would the main conclusions hold when derivation validity is checked and coarser method categories are used? What is the level of agreement among the human annotators?

Reviewer_Q3diOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper propose ACR, an inference-time method for MoE that assign global and private pool of experts among different samples for the same question to enhance diversity. The shared pool contains expert that all samples can use, the private pool contains experts that only that sample can use; it preserve the frozen router, while the router will only choose from the unmasked expert. On a self-built 126-problem maths and physics benchmark, ACR increase the number of problems that at least two distinct correct methods compared to best temperature sampling, while preserving most correctness. It also reports a 30 human audited problem subset.

Critical strengths
  1. It shows that simply restricting which experts a trajectory can use in a frozen MoE produces more distinct solution methods than temperature sampling and than routing-noise methods (RoE, Expert-Sample), at a small correctness cost (about 2 points under Sol and the human audit at P=160). This is a good empirical finding about MoE inference, without training nor prompt engineering.
Critical weaknesses
  1. Novelty is narrower than presented, and the closest prior work is missing. Dynamic Experts Search (Han et al., 2025) already uses a training-free, per-trajectory routing configuration that is kept fixed within a trajectory and varied across trajectories as an explicit source of diversity at test time. DES is a close prior work and a more relevant baseline than the per-token routing-noise controls (RoE, Expert-Sample). It is neither cited nor compared. Beyond the features that included in DES, ACR adds two new components: (a) the shared/private mask construction and (b) persistence of the mask across the trajectory. Table 7 shows that removing either component alone costs only a few problems (2.67 and 5.33 of 126, no significance test), while both controls still reach ~40 versus 27 for temperature. The baseline that would attribute this gain is missing: greedy decoding with a random expert subset per layer, redrawn at every token, with no shared core and no persistence. Without it, the paper cannot show that the shared/private, persistent construction, rather than masking experts at all, is what produces the diversity.

  2. The judgement from GPT and GLM is suspicious. GPT-5.6 Sol scores every trajectory from the temperature, Expert-Sample and RoE arms as correct (100.00% in Table 2), while GLM 5.3 gives 93.84–97.42 % for the temperature arms (Table 3). These problems have verifiable answers, so correctness should be close to objective. A disagreement of this size calls the judges into question. Specially, in Appendix A, the judge prompt is: "read raw predictions of jsons manually fully." and "in the end print pool-wise UCM@4 and number of problems with atleast 2 solutions and number of problems with atleast 3 solutions." This suggests that the judgment was not made problem by problem. Instead, the model appears to read all responses of an arm at once and output only aggregate metric. The prompt itself implies the judge had file access to prior artifacts, so blinding is unlikely.

Questions for authors
  1. Was the LLM judge invoked once per problem (4 trajectories) with programmatic aggregation, or over larger pools? Please report inter-judge agreement on per-trajectory correctness and per-problem UCM@4. Releasing the per-trajectory labels would also help.
  2. Why is T=1.1 absent from the Sol evaluation, and how do you explain Sol assigning exactly 100.00% correctness to four baseline arms while GLM assigns 93.8–99.8% to the same outputs?
  3. Please add two matched-budget controls: (i) a DES-style arm with a different number of activated experts per trajectory, and (ii) greedy decoding with a random expert subset masked per layer, redrawn every token, with no shared core or persistence.
1 comment

Thank you for requesting these controls and judging details. We completed both controls and compared GLM and Sol method counts.

W1/Q3(i): DES and activation-count control

"Please add ... a DES-style arm with a different number of activated experts per trajectory..."

DES varies how many experts are activated, carries that count along search paths, and uses verifier-guided selection. ACR keeps k=8k=8 and changes which experts each trajectory can access. We will cite DES; persistence alone is not our novelty claim.

Masking can skip higher-ranked experts without changing their count: with experts ranked 1>2>3>41>2>3>4, ACR can select {1,3}\{1,3\}; unrestricted top-kk cannot.

Our DES-style control generates one greedy trajectory for each k∈{6,7,9,10}k\in\{6,7,9,10\}, with all experts eligible and each count fixed across layers and tokens. It uses the same 126 questions, prompt, repetition penalty, output cap, Sol judge, and judging prompt as ACR.

ArmRunsN≥2N_{\geq2}N≥3N_{\geq3}CorrectnessUCM@4
ACR P1603 seeds45.004.0097.69%1.3915
DES-style1 sweep28.003.0099.80%1.2460

Counts are out of 126, averaged across ACR seeds. ACR yields 17 more two-method problems (60.7% more), with correctness lower by 2.11 points. This tests activation-count variation, not full DES. The control has no handoff; average k=8k=8 does not guarantee equal compute because output lengths differ.

For arXiv-only work, ICLR guidelines say missing comparisons alone do not justify rejection. This depends on publication status; such work still informs novelty.

W1/Q3(ii): Shared/private masks and persistence

"...a random expert subset per layer, redrawn at every token, with no shared core and no persistence."

The new control independently draws 184 eligible experts per trajectory and layer at every decoding step. All four arms use top-8 routing, greedy decoding, the same prompt, 2,000-token handoff, and output cap. The new arm covers all 126 questions over three seeds with the same Sol judge and instructions.

ConstructionPersistent?N≥2N_{\geq2}N≥3N_{\geq3}CorrectnessUCM@4
Shared/privateYes45.004.0097.69%1.3915
Shared/privateNo42.333.3399.21%1.3574
IndependentYes39.673.6796.49%1.3212
Independent, newNo32.332.3399.47%1.2778

Temperature T=0.9T=0.9 gives N≥2=27.00N_{\geq2}=27.00, correctness 100.00%, and UCM@4 1.2328.

ACR yields 12.67 more two-method problems than independent resampling (39.2% more), with correctness lower by 1.78 points. This contrast represents 70.4% of ACR's 18-problem gain over temperature. Persistence helps under both constructions; shared/private masks help with or without persistence. The closest-control gap remains 2.67 problems. These comparisons are descriptive, without paired significance tests or proof that either choice is necessary.

W2/Q1: Judging and agreement

"Was the LLM judge invoked once per problem ... or over larger pools? Please report inter-judge agreement..."

GLM and Sol ran through harnesses that spawned a subagent for each problem. Subagents read raw predictions in interactive file-reading sessions using the same instructions. Delegation alone does not establish blinding or isolation from other artifacts.

We matched available counts on identical generations: 126 questions across three seeds, or 378 paired judgments per condition.

ConditionExact UCM@4 agreementAgreement on ≥2\geq2Weighted κ\kappa [95% CI]
ACR P16076.19%77.78%0.392 [0.277,0.504]
Temperature T=0.991.01%91.80%0.701 [0.562,0.815]

Kappa uses linear weights. Intervals resample questions 20,000 times, keeping each question's seeds together. Lower ACR agreement shows evaluator sensitivity. These compare counts, not method identities, and do not reconstruct all submitted aggregates. One inconsistent GLM record per condition is retained; excluding it gives κ=0.387\kappa=0.387 and 0.6980.698.

The submitted blinded 30-problem human audit provides another check with model and setting identities hidden: N≥2=12N_{\geq2}=12 versus 99 on Qwen and 1414 versus 77 on Gemma (ACR versus temperature). We will release all available labels.

Q2: Correctness discrepancy and T=1.1 results

"Why is T=1.1 absent ... Sol assigning exactly 100.00% correctness..."

Expert-Sample: GLM accepts 503/504; Sol, 504/504 (99.80% agreement). They differ on the Wallis product ∏n=1∞4n24n2−1=π/2\prod_{n=1}^{\infty}\frac{4n^2}{4n^2-1}=\pi/2: GLM rejects the unfinished ending; Sol credits the earlier valid Stirling proof. This soft-correctness boundary also explains temperature-arm gaps.

For T=1.1T=1.1, Sol's three-seed means are N≥2=29.00N_{\geq2}=29.00, N≥3=2.33N_{\geq3}=2.33 (out of 126), Correctness =99.60%=99.60\% and UCM@4 =1.253=1.253.

Reviewer_kJyaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

How can an MoE model generate rollouts that follow genuinely different solution methods rather than converge on the same approach? The authors find that increasing the sampling temperature alone does not produce much method diversity, and investigate an alternative way to address this problem.

The authors propose Agent-Conditioned Routing (ACR). Simply put, this method randomly disables a subset of experts for each reasoning trajectory, allowing routing only to that trajectory's shared and private experts. This encourages the rollouts to follow more diverse solution methods.

This method produces more diverse rollouts than adjusting the sampling temperature. For example, under GLM evaluation, ACR achieves N≥2 = 37.3, compared with 18.7 for the strongest temperature setting (T = 0.9). However, it performs worse than DivPrompt, which provides previous solutions and asks the model to use a different method.

The evaluation uses 126 math and physics problems screened from 200 manually written problems. A problem is retained if six temperature-sampled rollouts show multiple solution methods and include at least one correct solution. The main metric, N≥2, counts problems for which the four rollouts contain at least two distinct primary methods among soft-correct solutions, as assessed by two model judges, GLM 5.3 and GPT-5.6 Sol.

Critical strengths
  1. The method is simple. It only requires masking routing logits separately for each trajectory, adds no parameters, and keeps the number of active experts unchanged. It can be batched like ordinary sampling.

  2. The research question is interesting.

  3. The paper openly reports unfavorable results. The authors acknowledge that DivPrompt substantially outperforms ACR, that correctness drops at P = 96 and on Gemma, that eligibility-matched controls retain most of the effect, and that the handoff analysis has limitations. This transparency is appreciated.

Critical weaknesses

W1. The improvement is modest, and ACR is outperformed by DivPrompt.

  • The absolute improvement is modest. Under Sol evaluation, mean UCM@4 increases from 1.2328 to 1.3915. ACR still yields fewer than two distinct correct methods on approximately 64% of the problems (126 − 45 = 81).
  • DivPrompt, a simple prompting strategy that can be automated, reaches 125/126, substantially outperforming ACR.

W2. The benchmark has a limited scope, and the practical value of diversity is not demonstrated.

  • The 126 test problems mostly involve classical textbook-level derivations, such as inclusion–exclusion, rolling down an incline, and the Basel series. Correctness is generally near ceiling (94–100%). These problems were specifically screened for evidence of multiple solution methods, and the development set was AI-generated. The benchmark's availability, difficulty, and potential data contamination are also unclear.
  • The paper does not demonstrate the practical value of method diversity. It does not evaluate pass@k or coverage on harder benchmarks such as AIME, HMMT, Putnam, or OlympiadBench, nor does it test whether diversity improves majority voting or verifier-based selection. Appendix D provides only a single Putnam example, which is insufficient evidence.
  • ACR reduces correctness in several settings: at P = 96, correctness is 52.78% under GLM and 60.71% under Sol; on Gemma, it is 84.9% versus 87.6%; and in the Gemma human audit, it is 86.2% versus 100%. A utility analysis is needed to show that the additional diversity is worth these costs.
Questions for authors

see weakness

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.