On-Policy Co-Distillation (OPCoD)

• OPCoD is a training pattern that leverages on-policy rollouts and multi-source supervision to integrate feedback from peers and specialized teachers.
• It enables bidirectional peer tutoring and multi-teacher integration, yielding mutual performance enhancements while avoiding negative transfer.
• Empirical studies report significant metric improvements and efficiency gains, illustrating OPCoD’s effectiveness across diverse, multi-domain tasks.

On-Policy Co-Distillation (OPCoD) denotes an on-policy distillation pattern in which a learner is trained on states induced by its own policy, while supervisory signals are aggregated across multiple sources rather than supplied by a single fixed teacher. In recent 2026 literature, the label is used non-uniformly: it can denote bidirectional peer tutoring between LLMs, multi-teacher integration of specialized diffusion teachers, dual supervision from a privileged teacher and a privileged student branch, or broader multi-policy OPD mechanisms in which peers, specialists, or co-evolving experts provide dense feedback on student rollouts. One paper also states explicitly that “On-Policy Co-Distillation (OPCoD)” is a naming confusion relative to “On-Policy Context Distillation (OPCD),” indicating that the terminology is not yet standardized (Byeon et al., 12 Jun 2026, Li et al., 14 May 2026, Ye et al., 12 Feb 2026).

1. Terminology, scope, and relation to on-policy distillation

Under a formula-driven OPD taxonomy, OPCoD can be instantiated as a multi-policy on-policy distillation mechanism: the active student policy induces contexts $h_t$, and one or more peer or teacher policies provide feedback distributions on those same contexts (Zhang, 22 Jun 2026). AsyncOPD states the same generalization operationally, describing OPCoD as an extension of OPD to multiple teachers and/or multiple students, including mutual co-distillation (Kang et al., 23 Jun 2026). In this sense, OPCoD is not a single algorithmic recipe but a structural training pattern defined by on-policy state generation plus multi-source supervision.

This broad usage coexists with narrower, paper-specific definitions. “Be My Tutor” uses OPCoD for a bidirectional peer-feedback framework in which two models, each stronger in a different domain, co-evolve by tutoring each other through on-policy feedback and target mutual Pareto improvement (Byeon et al., 12 Jun 2026). “Co-Evolving Policy Distillation” presents CoPD as bidirectional on-policy distillation among concurrently trained expert branches; its accompanying synthesis identifies CoPD functionally as OPCoD (Gu et al., 29 Apr 2026). “DiffusionOPD” states that its multi-task distillation procedure is precisely an on-policy co-distillation procedure for diffusion-based text-to-image generation, because multiple task-specialized teachers are distilled into one student on the student’s own denoising trajectories (Li et al., 14 May 2026).

A terminological caveat is necessary. “On-Policy Context Distillation for LLMs” defines OPCD, not OPCoD, and explicitly states that no separate “co-distillation” variant is defined there; in that paper, the exact mechanism is on-policy distillation against a context-conditioned teacher for internalizing in-context knowledge (Ye et al., 12 Feb 2026). This makes “OPCoD” best understood as a non-uniform label whose meaning depends on the specific paper.

2. Canonical formulations

A generic OPCoD formulation treats the active learner $i$ as the owner of the rollout distribution and constructs a per-state mixture of supervising policies: $$q_{\mathrm{mix},i}(\cdot \mid h_t)=\sum_{j\in T(i)} \alpha_{ij,t}\,q_j(\cdot \mid h_t),$$ where $h_t=(x,y_{<t})$ is induced by the current student, $T(i)$ is the relevant peer/teacher set, and $\alpha_{ij,t}$ are mixture weights. The survey gives a direct distributional route,

$$L_{\mathrm{direct},i} = \mathbb{E}_{h_t\sim d_i}\sum_t w_t\, D_i\!\big[\pi^i_\theta(\cdot\mid h_t), q_{\mathrm{mix},i}(\cdot\mid h_t); \Omega_{i,t}\big],$$

together with a policy-gradient-style route built from the immediate log-ratio reward

$$r^i_t=\log q_{\mathrm{mix},i}(y_t\mid h_t)-\log \pi^i_\theta(y_t\mid h_t).$$

Temporal credit can then use immediate estimates, return-to-go, discounted returns, baseline-corrected estimators, or GAE-OPD; vocabulary routing can be separated from temporal credit through explicit support construction and routed losses such as CR-OPD (Zhang, 22 Jun 2026).

The peer-feedback OPCoD formulation in “Be My Tutor” is a concrete discrete-token instance. For tutee $i$, an on-policy rollout $y\sim \pi^i_S(\cdot\mid x)$ is supervised by a self-teacher conditioned on privileged context $i$0, where $i$1 is the earliest verified correct rollout and $i$2 is peer feedback: $i$3 Feedback is admitted only when the tutor passes cognizance-based gating, $i$4, with $i$5 in the reported experiments (Byeon et al., 12 Jun 2026).

DiffusionOPD extends the same logic from discrete tokens to continuous-state Markov processes. A diffusion sampler defines a latent trajectory $i$6, the student rolls out the denoising path, and a task-specific teacher is queried at each visited latent to supply the one-step transition distribution. Under shared isotropic covariance, the per-step KL reduces to mean matching,

$i$7

and the deterministic ODE specialization becomes a pointwise $i$8 objective (Li et al., 14 May 2026).

3. Major realizations in 2026 literature

Realization	Supervision sources	Distinctive mechanism
Peer-feedback OPCoD	Two co-evolving students	Verified correct rollout, cognizance-based gating, feedback anchoring
DiffusionOPD as OPCoD	Multiple task-specialized teachers	Per-step reverse KL or deterministic $i$9 on student denoising trajectories
MAD-OPD	Debate-driven teacher collective	Post-debate confidence weighting, task-adaptive divergence
DOPD	Privileged teacher + privileged student anchor	Advantage-aware token routing across LH/LL/HT/HS regimes
CoDistill-GRPO	Large and small policies	On-policy KD reward for small model, importance-reweighted large-model update
CoPD	Concurrent expert branches	Interleaved RLVR and bidirectional OPD during co-evolution
FA-OPD	Co-trained FM teacher	Dual reward and action channels on student rollouts

“Be My Tutor” defines the most literal peer-to-peer form: each model alternates between tutor and tutee roles, the tutee samples $$q_{\mathrm{mix},i}(\cdot \mid h_t)=\sum_{j\in T(i)} \alpha_{ij,t}\,q_j(\cdot \mid h_t),$$0 on-policy rollouts, the earliest correct rollout is used as the self-anchor, and the tutor provides anchored feedback only if it is sufficiently cognizant relative to the best validation performer across domains (Byeon et al., 12 Jun 2026). The framework is explicitly designed for mutual Pareto improvement rather than one-way transfer.

DiffusionOPD realizes co-distillation in a many-teacher, one-student regime. Teachers are trained independently on domains such as OCR, compositional alignment, and aesthetics, then integrated by alternating student rollouts across tasks and aggregating per-step distillation losses over a complete round. The paper emphasizes that this decouples single-task exploration from multi-task integration and replaces sparse terminal rewards with dense distributional supervision (Li et al., 14 May 2026).

MAD-OPD pushes the multi-teacher view further by replacing a single teacher with a deliberative collective. Two teachers debate over the student’s on-policy state for $$q_{\mathrm{mix},i}(\cdot \mid h_t)=\sum_{j\in T(i)} \alpha_{ij,t}\,q_j(\cdot \mid h_t),$$1 rounds, self-report scalar confidences $$q_{\mathrm{mix},i}(\cdot \mid h_t)=\sum_{j\in T(i)} \alpha_{ij,t}\,q_j(\cdot \mid h_t),$$2, and produce a confidence-weighted mixture. The paper pairs this with OPAD for multi-step agentic tasks and adopts $$q_{\mathrm{mix},i}(\cdot \mid h_t)=\sum_{j\in T(i)} \alpha_{ij,t}\,q_j(\cdot \mid h_t),$$3 for agentic stability and reverse KL for code generation (Wang et al., 2 May 2026).

DOPD presents another OPCoD interpretation: token-level supervision is dynamically routed between a privileged teacher policy and a privileged student policy based on a privilege advantage gap and relative probabilities. Its key concern is “privilege illusion,” the conflation of transferable capability with information asymmetry; hard routing separates low-gap/high-confidence, low-gap/low-confidence, high-gap/teacher-confident, and high-gap/student-confident tokens into LH, LL, HT, and HS regimes (Yu et al., 29 Jun 2026).

CoDistill-GRPO instantiates co-distillation inside GRPO. The small model receives a dense on-policy KD reward built from token-level log-probability ratios against the large model on the small model’s own rollouts, while the large model is trained on small-model rollouts with per-token importance reweighting (Kwon et al., 9 May 2026). CoPD applies the same mutual idea to expert branches trained on distinct capabilities: branches continue native RLVR while performing bidirectional OPD on one another’s domains, keeping behavioral overlap high enough for effective absorption (Gu et al., 29 Apr 2026). FA-OPD transposes co-distillation to embodied control by letting a co-trained flow-matching teacher supply both an expert-likeness reward channel and dense action targets on student-visited states (Wan et al., 26 May 2026).

4. Optimization mechanisms and design choices

Across these realizations, the defining invariant is that supervision is evaluated on the learner’s own state distribution. In language-model settings this means force-decoding teacher or peer distributions on the active student’s prefixes; in diffusion models it means teacher transition kernels on student-visited latents; in embodied control it means reward or action targets on student-visited state–action pairs (Kang et al., 23 Jun 2026, Li et al., 14 May 2026, Wan et al., 26 May 2026).

The choice of divergence is treated as a substantive design variable rather than an implementation detail. MAD-OPD derives a task-adaptive rule: $$q_{\mathrm{mix},i}(\cdot \mid h_t)=\sum_{j\in T(i)} \alpha_{ij,t}\,q_j(\cdot \mid h_t),$$4 for agentic tasks because its per-token logit gradients are uniformly bounded, and reverse KL for code because mode concentration avoids incoherent splicing of distinct implementations (Wang et al., 2 May 2026). AsyncOPD shows that KL direction also matters under stale-policy data: teacher-weighted forward KL is more robust to staleness, whereas student-weighted reverse KL is vulnerable because current-student support can fall outside the stale teacher-score cache. For reverse-KL OPCoD under asynchrony, the paper recommends recomputing the advantage $$q_{\mathrm{mix},i}(\cdot \mid h_t)=\sum_{j\in T(i)} \alpha_{ij,t}\,q_j(\cdot \mid h_t),$$5 at learner time and using unclipped old-to-current importance sampling, together with multi-sample Monte Carlo caches (Kang et al., 23 Jun 2026).

Support construction and routing are likewise central. DOPD uses full-vocabulary $$q_{\mathrm{mix},i}(\cdot \mid h_t)=\sum_{j\in T(i)} \alpha_{ij,t}\,q_j(\cdot \mid h_t),$$6 on HT tokens, Top-$$q_{\mathrm{mix},i}(\cdot \mid h_t)=\sum_{j\in T(i)} \alpha_{ij,t}\,q_j(\cdot \mid h_t),$$7 reverse KL against the teacher on LH tokens, and Top-$$q_{\mathrm{mix},i}(\cdot \mid h_t)=\sum_{j\in T(i)} \alpha_{ij,t}\,q_j(\cdot \mid h_t),$$8 reverse KL against a privileged student anchor on LL and HS tokens, with $$q_{\mathrm{mix},i}(\cdot \mid h_t)=\sum_{j\in T(i)} \alpha_{ij,t}\,q_j(\cdot \mid h_t),$$9, $h_t=(x,y_{<t})$0, and $h_t=(x,y_{<t})$1 in the reported implementation (Yu et al., 29 Jun 2026). “Be My Tutor” instead controls supervision through verification and natural-language gates: exact answer matching defines whether a self-anchor exists, cognizance-based gating silences incognizant peers, and feedback anchoring requires a tutor to extract an exact question concept in tags before producing a 3–5 sentence critique (Byeon et al., 12 Jun 2026).

Scheduling varies with the application. DiffusionOPD alternates tasks in round-robin order, sets the gradient accumulation factor to the number of tasks $h_t=(x,y_{<t})$2, averages losses over a complete sweep, and updates once per round (Li et al., 14 May 2026). CoPD alternates branch-specific RLVR with mutual OPD in parallel branches and finds that an $h_t=(x,y_{<t})$3 ratio of $h_t=(x,y_{<t})$4 gives the best overall accuracy in ablations (Gu et al., 29 Apr 2026). CoDistill-GRPO uses a different co-schedule: the large model may supply $h_t=(x,y_{<t})$5 hint tokens only in the first epoch, the small model generates $h_t=(x,y_{<t})$6 rollouts per prompt, and top-and-bottom $h_t=(x,y_{<t})$7 rollouts are downsampled using the KD-augmented score $h_t=(x,y_{<t})$8 (Kwon et al., 9 May 2026).

Teacher dynamics differ sharply across papers. Frozen teacher-student OPCD is reported as more stable than self-distillation in the context-distillation paper (Ye et al., 12 Feb 2026). MAD-OPD freezes teachers but changes their aggregate supervision through debate (Wang et al., 2 May 2026). DOPD freezes the privileged teacher and uses a stop-gradient privileged student anchor (Yu et al., 29 Jun 2026). CoPD and FA-OPD are genuinely co-evolving: teachers or expert branches change during training, which increases adaptation but also introduces non-stationarity (Gu et al., 29 Apr 2026, Wan et al., 26 May 2026).

5. Reported empirical behavior

The strongest fully explicit use of the term OPCoD appears in “Be My Tutor,” where the method achieves mutual Pareto improvement across all evaluated domain pairs and students on Science QA-style tasks. Representative avg@16 results include Materials–Physics (Student 1), where the average rises from $h_t=(x,y_{<t})$9 initially to $T(i)$0 under OPCoD; Chemistry–Materials (Student 1), from $T(i)$1 to $T(i)$2; and Physics–Chemistry (Student 2), from $T(i)$3 to $T(i)$4. Solo SDPO can degrade a model’s native domain, whereas OPCoD is reported to avoid that negative transfer (Byeon et al., 12 Jun 2026).

DiffusionOPD reports that its OPCoD-style multi-task distillation surpasses both multi-reward RL and cascade RL baselines in training efficiency and final quality. The average normalized score is $T(i)$5 for DiffusionOPD, versus $T(i)$6 for Cascade NFT, $T(i)$7 for Multi-Task GRPO-Guard, and $T(i)$8 for Multi-Task NFT. Reported representative scores for DiffusionOPD are GenEval $T(i)$9, OCR $\alpha_{ij,t}$0, PickScore $\alpha_{ij,t}$1, ClipScore $\alpha_{ij,t}$2, HPSv2.1 $\alpha_{ij,t}$3, Aesthetics $\alpha_{ij,t}$4, ImageReward $\alpha_{ij,t}$5, and UniReward $\alpha_{ij,t}$6. Student distillation adds approximately $\alpha_{ij,t}$7 hours on top of the longest teacher ($\alpha_{ij,t}$8 h), whereas multi-task RL baselines exceed $\alpha_{ij,t}$9–$$L_{\mathrm{direct},i} = \mathbb{E}_{h_t\sim d_i}\sum_t w_t\, D_i\!\big[\pi^i_\theta(\cdot\mid h_t), q_{\mathrm{mix},i}(\cdot\mid h_t); \Omega_{i,t}\big],$$0 h; the ODE regime can be up to five times more efficient than SDE at $$L_{\mathrm{direct},i} = \mathbb{E}_{h_t\sim d_i}\sum_t w_t\, D_i\!\big[\pi^i_\theta(\cdot\mid h_t), q_{\mathrm{mix},i}(\cdot\mid h_t); \Omega_{i,t}\big],$$1 (Li et al., 14 May 2026).

MAD-OPD reports first place across all six teacher–student configurations it evaluates. In the Qwen3 $$L_{\mathrm{direct},i} = \mathbb{E}_{h_t\sim d_i}\sum_t w_t\, D_i\!\big[\pi^i_\theta(\cdot\mid h_t), q_{\mathrm{mix},i}(\cdot\mid h_t); \Omega_{i,t}\big],$$2B+$$L_{\mathrm{direct},i} = \mathbb{E}_{h_t\sim d_i}\sum_t w_t\, D_i\!\big[\pi^i_\theta(\cdot\mid h_t), q_{\mathrm{mix},i}(\cdot\mid h_t); \Omega_{i,t}\big],$$3B$$L_{\mathrm{direct},i} = \mathbb{E}_{h_t\sim d_i}\sum_t w_t\, D_i\!\big[\pi^i_\theta(\cdot\mid h_t), q_{\mathrm{mix},i}(\cdot\mid h_t); \Omega_{i,t}\big],$$4B setting, the agentic average rises from $$L_{\mathrm{direct},i} = \mathbb{E}_{h_t\sim d_i}\sum_t w_t\, D_i\!\big[\pi^i_\theta(\cdot\mid h_t), q_{\mathrm{mix},i}(\cdot\mid h_t); \Omega_{i,t}\big],$$5 under single-teacher OPD to $$L_{\mathrm{direct},i} = \mathbb{E}_{h_t\sim d_i}\sum_t w_t\, D_i\!\big[\pi^i_\theta(\cdot\mid h_t), q_{\mathrm{mix},i}(\cdot\mid h_t); \Omega_{i,t}\big],$$6 under MAD-OPD, and the code average rises from $$L_{\mathrm{direct},i} = \mathbb{E}_{h_t\sim d_i}\sum_t w_t\, D_i\!\big[\pi^i_\theta(\cdot\mid h_t), q_{\mathrm{mix},i}(\cdot\mid h_t); \Omega_{i,t}\big],$$7 to $$L_{\mathrm{direct},i} = \mathbb{E}_{h_t\sim d_i}\sum_t w_t\, D_i\!\big[\pi^i_\theta(\cdot\mid h_t), q_{\mathrm{mix},i}(\cdot\mid h_t); \Omega_{i,t}\big],$$8. Debate adds substantial code gains over MT-OPD, and confidence weighting adds $$L_{\mathrm{direct},i} = \mathbb{E}_{h_t\sim d_i}\sum_t w_t\, D_i\!\big[\pi^i_\theta(\cdot\mid h_t), q_{\mathrm{mix},i}(\cdot\mid h_t); \Omega_{i,t}\big],$$9 Ag-Avg in ablations (Wang et al., 2 May 2026).

DOPD reports consistent gains over Vanilla OPD and other counterparts in both LLM and VLM settings. For Qwen3-$$r^i_t=\log q_{\mathrm{mix},i}(y_t\mid h_t)-\log \pi^i_\theta(y_t\mid h_t).$$0B$$r^i_t=\log q_{\mathrm{mix},i}(y_t\mid h_t)-\log \pi^i_\theta(y_t\mid h_t).$$1B, the student average is $$r^i_t=\log q_{\mathrm{mix},i}(y_t\mid h_t)-\log \pi^i_\theta(y_t\mid h_t).$$2, the teacher average $$r^i_t=\log q_{\mathrm{mix},i}(y_t\mid h_t)-\log \pi^i_\theta(y_t\mid h_t).$$3, Vanilla OPD $$r^i_t=\log q_{\mathrm{mix},i}(y_t\mid h_t)-\log \pi^i_\theta(y_t\mid h_t).$$4, and DOPD $$r^i_t=\log q_{\mathrm{mix},i}(y_t\mid h_t)-\log \pi^i_\theta(y_t\mid h_t).$$5, which the paper summarizes as $$r^i_t=\log q_{\mathrm{mix},i}(y_t\mid h_t)-\log \pi^i_\theta(y_t\mid h_t).$$6 gap recovery. For Qwen3-VL-$$r^i_t=\log q_{\mathrm{mix},i}(y_t\mid h_t)-\log \pi^i_\theta(y_t\mid h_t).$$7B$$r^i_t=\log q_{\mathrm{mix},i}(y_t\mid h_t)-\log \pi^i_\theta(y_t\mid h_t).$$8B, the student average is $$r^i_t=\log q_{\mathrm{mix},i}(y_t\mid h_t)-\log \pi^i_\theta(y_t\mid h_t).$$9, the teacher $i$0, Vanilla OPD $i$1, and DOPD $i$2, corresponding to $i$3 gap recovery (Yu et al., 29 Jun 2026).

CoDistill-GRPO shows that co-distillation can improve small-model RLVR under sparse rewards while reducing rollout cost for the large model. On Minerva, Qwen2.5-Math-$i$4B reaches $i$5 with CoDistill-GRPO, versus $i$6 for GRPO and $i$7 for the base model. The paper also reports an approximate $i$8 speedup for the larger model when rollout generation is done solely by the small model (Kwon et al., 9 May 2026). CoPD reports that its co-evolving all-in-one model reaches an overall average of $i$9 in the two-branch setting, compared with $y\sim \pi^i_S(\cdot\mid x)$0 for Mixed RLVR, $y\sim \pi^i_S(\cdot\mid x)$1 for OPD (V→T), and $y\sim \pi^i_S(\cdot\mid x)$2 for OPD (T→V), and $y\sim \pi^i_S(\cdot\mid x)$3 in the three-branch setting versus $y\sim \pi^i_S(\cdot\mid x)$4 for MOPD and $y\sim \pi^i_S(\cdot\mid x)$5 for Mixed RLVR (Gu et al., 29 Apr 2026).

Systems results show that asynchronous OPCoD-style training can materially improve throughput without major loss of accuracy. AsyncOPD reports $y\sim \pi^i_S(\cdot\mid x)$6 to $y\sim \pi^i_S(\cdot\mid x)$7 throughput improvements over strict synchronous training while reaching comparable accuracy, and attributes the remaining difficulty chiefly to stale-policy effects in reverse-KL estimators and finite teacher-score caches (Kang et al., 23 Jun 2026). In embodied control, FA-OPD reports best final performance in $y\sim \pi^i_S(\cdot\mid x)$8 tasks, including Hand-rotate $y\sim \pi^i_S(\cdot\mid x)$9, Fetch-pick $i$00, Maze2d $i$01, Hopper $i$02, and Walker2d $i$03, while showing stronger robustness under noisy or limited demonstrations (Wan et al., 26 May 2026).

6. Limitations, failure modes, and research directions

A first limitation is conceptual: the term itself is unsettled. The literature uses OPCoD for peer tutoring, multi-teacher mixture distillation, dual privileged supervision, co-evolving expert branches, and GRPO co-training, while one paper explicitly rejects the term for its own context-distillation setting (Ye et al., 12 Feb 2026). This suggests that current usage is umbrella-like rather than canonical.

Several failure modes recur across implementations. Teacher quality remains a hard ceiling in multi-teacher diffusion distillation: distillation cannot exceed teacher capabilities, and behaviors such as overfitting, reward hacking, or poor generalization may transfer into the unified student (Li et al., 14 May 2026). Single-teacher OPD is described by MAD-OPD as capped by the teacher’s capability ceiling, motivating deliberative collectives rather than naive one-teacher supervision (Wang et al., 2 May 2026). DOPD adds a distinct failure mode, “privilege illusion,” in which extra training-time information induces shortcuts that cannot be replicated at inference time (Yu et al., 29 Jun 2026).

Stability problems are equally prominent. OPCD reports that self-distillation can be unstable and substantially underperform frozen teacher-student configurations (Ye et al., 12 Feb 2026). AsyncOPD shows that stale-policy data degrade reverse-KL OPD more sharply than forward-KL variants, and that stale sparse top-$i$04 caches are biased relative to the current objective when current student support shifts outside the cached actions (Kang et al., 23 Jun 2026). CoPD identifies a separate instability trade-off: too much OPD risks behavioral collapse because teacher and student become indistinguishable, while too little OPD yields large behavioral gaps and low absorption efficiency (Gu et al., 29 Apr 2026).

Task conflict and scaling remain open problems. DiffusionOPD notes that some capabilities may inherently conflict in the student parameterization, and identifies adaptive weighting, curriculum, or gating as possible mitigations (Li et al., 14 May 2026). “Be My Tutor” studies only pairwise co-distillation and notes that extending to more than two peers is unexplored; it proposes graph-based gating or tournament-style tutoring as future directions (Byeon et al., 12 Jun 2026). AsyncOPD points to unresolved systems questions for multi-teacher aggregation under bandwidth constraints, teacher subsampling with unbiased aggregation, moving-teacher mutual distillation, and scaling beyond a single 8-GPU node (Kang et al., 23 Jun 2026).

The forward agenda is correspondingly broad. The OPD survey maps natural extensions for OPCoD to multi-teacher mixtures, co-evolving peer teachers, GAE-style temporal credit, CR-OPD vocabulary routing, verifier-weighted gates, and trust-region stabilization (Zhang, 22 Jun 2026). Specific papers add more specialized directions: dynamic task scheduling, uncertainty-aware teacher selection, and DMD/TDM-style objective mixing for diffusion co-distillation (Li et al., 14 May 2026); multi-context aggregation and curriculum strategies for context internalization (Ye et al., 12 Feb 2026); improved confidence signals beyond self-reported debate confidence for multi-teacher supervision (Wang et al., 2 May 2026); and stronger verification for correctness and feedback quality in peer-tutoring OPCoD (Byeon et al., 12 Jun 2026). Collectively, these directions indicate that OPCoD is evolving less as a single method than as a design space for dense, on-policy, multi-source supervision.