Multi-domain capabilities are critical for practically applicable Large Reasoning Models (LRMs). Therefore, Multi-teacher On-Policy Distillation (MOPD) has become a crucial part of the post-training process, integrating the capabilities of multiple domain-specific teacher models into a single student model. However, the student usually fails to acquire multi-domain capabilities simultaneously due to undesirable domain heterogeneity. Existing works have address this problem in view of temporal uneven efficiency and spatial hard allocation in external multi-teacher supervision. While in broader multi-domain training, varying degrees of performance degradation is still largely unexplained. In this paper, we shift the focus on how the student learns and integrates multi-domain capabilities internally, and then propose Janus-MOPD, a general learning dynamics framework of geometric representations for LRMs. Specifically, our main focus is on the internal performance of multi-domain capabilities on both temporal and spatial dimensions within the student. Temporally, our study shows that the student exhibits a Expansion-Consolidation-Optimization (ECO) three-phase learning dynamics, together with inherent limitations in learning efficiency and scaling. Spatially, we summarize two types of subspace locking within and across domains by tracing the representation space of the student, while explaining its clear learning directions but inherent cross-domain conflict. Therefore, Janus-MOPD yields a unified learning dynamics formulation, a geometric tracing methodology, and systematic findings of how the student learns and integrates multi-domain capabilities. Overall, we provide a Janus perspective to internally trace the continual evolution of the multi-domain capabilities within the model itself.
Figures & tables
Figure 2: Representation geometry across post-training methods. (a) Mixed-RL; (b) Standard-MOPD; (c) Open-MOPD; (d) MOPD-Router. Each sector represents one domain: solid and dashed curves show eRank and αReQ , respectively, and inner heatmaps show Sim16 .
Figure 3: Benchmark accuracy trajectories across five domains. (a) Qwen3-4B-Instruct (16K); (b) Safin1-4B (32K). Each domain score averages its benchmarks, following Table 1 .
Figure 4: Domain-specific temporal dynamics. (a) eRank ; (b) αReQ . Outlined symbols mark eRank maxima and spectral-decay minima. Gray bands show the terminal P10–P90 range of single-domain OPD. Red horizontal guides mark its eRank maximum, red vertical guides mark its stabilization steps.
Figure 4
Figure 6: Within-domain subspace persistence with two temporal references. (a) The current domain versus its previous saved anchor; (b) the current domain versus its step-0 subspace.
Figure 7: Cross-domain subspace overlap with two temporal references. For each domain k at step t , (a) averages Sim16 with the other four domains at tprev , and (b) with those domains at step 0.
Figure 8: Spatial responses to domain interventions. Rows show Sim16 probe previous and initial subspaces within each domain and across domains.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Role
Model
Setting
Main student
Qwen3-4B-Instruct-2507
16K response limit
Teacher pool
Five RL experts from the same Qwen3 backbone
One expert per domain
Single-domain
Five independently distilled Qwen3 students
One domain per student
Additional backbone
Safin1-4B
32K response limit
Appendix
Table 3: Models and their roles. Response limits describe capability evaluation; the shared Qwen3 training configuration appears in Table 5 .
Domain
Training
Evaluation
Math
NuminaMath-1.5 (10,000)
AIME24 / AIME25 / AIME26 (30 each)
Science
Nemotron-RL-knowledge-mcqa (10,000)
SciBench (580); GPQA-Diamond (198)
Code
TACO-derived (8,460); LeetCode (1,540)
LiveCodeBench v5 (167) / v6 (175)
ToolUse
ToolACE (10,000)
ToolCall (496)
SafeSci
SafeSciTrain (10,000)
SafeSciBench / SciSafety (500)
Appendix
Table 4: Training and evaluation data across five domains.
Hyperparameter
Qwen3-4B-Instruct
Safin1-4B
General settings
Student / teacher architecture
Qwen3-4B-Instruct-2507
Safin1-4B
Configured updates / data seed
500 / 42
500 target / 43
Prompt mixture
10,000 records per domain
10,000 records per domain
Global rollout batch / optimization minibatch
1,024 / 256
1,024 / 256
Responses per prompt / optimization passes
1 / 1
1 / 1
Appendix
Table 5: Training configuration for Qwen3-4B and Safin1-4B with MOPD methods.
Method
Control
Value
Standard-MOPD
Prompt shares; teacher assignment
0.2 each; matched expert
Open-MOPD
Target contribution shares
0.2 each
Gap exponent; multiplier clipping
1.0 ; [0.05,20]
Initial-gap anchoring; advantage refresh
Enabled; enabled
MOPD-Router
Teacher aggregation; candidate count
Positive cosine; 16
Alignment tolerance
10−6
Appendix
Table 6: Method-specific settings of MOPD.
Figure 9: Learning-rate sensitivity of temporal geometry. Five-domain Standard-MOPD trajectories show (a) eRank and (b) αReQ ; outlined markers indicate maxima and minima, respectively. The default rate is 1.5×10−6 .
Figure 10: Learning-rate sensitivity of spatial geometry. (a) Within-domain Sim16 between consecutive saved anchors; (b) mean cross-domain Sim16 at the same anchor. Learning rates match Figure 9 .
Figure 11: Five-domain supervision coordination in Open-MOPD. (a) Realized token shares (solid) and prompt shares (dashed); (b) normalized domain loss coefficients; (c) mean absolute distillation signals on logarithmic vertical axes.
Figure 12: Five-expert coordination in MOPD-Router. (a) Mean teacher weight; (b) eligible positive-alignment rate; (c) mean specialization norm on the student top-16 candidate set. Columns denote teachers; statistics aggregate valid tokens across all prompt domains.
Method
Ks
Within min.
Within median
Cross-domain range
Standard-MOPD
8
0.9818
0.9972
[0.2002,0.3685]
12
0.9664
0.9965
[0.2417,0.3532]
16
0.9469
0.9966
[0.2378,0.3322]
Open-MOPD
8
0.9850
0.9976
[0.2006,0.3768]
12
0.9706
0.9968
[0.2442,0.3580]
16
0.9585
0.9962
[0.2420,0.3339]
Appendix
Table 7: Subspace-dimension sensitivity on fixed Qwen3 probes at t≥200 .
SIA-Lab of Tsinghua AIR and ByteDance Seed · Institute for AI Industry Research (AIR), Tsinghua University · Department of Computer Science and Technology, Tsinghua University