Skip to content

Academia Advances in Multimodal RL — Late 2025 & 2026

Scope: algorithmic papers proposing new training objectives or sampling strategies for RL fine-tuning of image/video diffusion and flow models. Robotics, language-only, and discrete-diffusion papers excluded. Sorted chronologically within each paradigm group. Math follows the same conventions as the main survey pages.

Papers covered elsewhere (not in this long tail): FlowGRPO, DanceGRPO, MixGRPO, CPS, GRPO-Guard, FlowDPPO, UniGRPO, SRPO, DiffusionNFT, AWM, and DGPO have dedicated pages; DDPO and Diffusion-DPO are covered as precursors in INDEX.md.


Master Index (sorted by date)

Date arXiv Paper Paradigm Key Problem
2025-09 2509.06040 BranchGRPO Policy Gradient Inefficient sequential rollouts; sparse credit assignment
2025-10 2510.00502 DAV Direct Preference Mode collapse under standard KL-RL; no EM framework
2025-10 2510.02692 P-GRAFT Direct Preference RAFT methods lack unified theory; terminal-only shaping suboptimal
2025-10 2510.18072 AC-Flow Direct Preference No actor-critic framework for continuous-time flow matching
2025-11 2511.19356 TaRoS Policy Gradient Goodhart's Law / reward saturation in video GRPO
2025-12 2512.04332 DDRL Direct Preference On-policy KL regularisation causes reward hacking at scale
2025-12 2512.04559 SQDF Direct Preference No principled Q-function for diffusion denoising chain
2025-12 2512.08153 TreeGRPO Policy Gradient Trajectory-level uniform advantages; high sampling cost
2025-12 2512.15347 Pro-GRPO Policy Gradient Large groups costly; reward-clustered samples give no signal
2025-12 2512.17951 SuperFlow Policy Gradient Fixed group size ignores per-prompt variance; coarse step advantages
2025-12 2512.21514 DiverseGRPO Policy Gradient GRPO collapses toward few high-reward modes
2025-12 2512.24138 GARDO Direct Preference KL regularisation impedes exploration; no uncertainty-aware gating
2026-01 2601.00423 E-GRPO Policy Gradient Low-entropy SDE steps waste compute; exploration budget uniform
2026-01 2601.04153 Diffusion-DRF Direct Preference Non-differentiable rewards block VLM-based feedback
2026-01 2601.05729 TAGRPO Policy Gradient Standard GRPO fails on I2V; needs relational trajectory alignment
2026-01 2601.12401 DRIFT Policy Gradient RL fine-tuning provably converges to Dirac delta (mode collapse)
2026-01 2601.20218 DenseGRPO Policy Gradient Terminal reward applied uniformly to all steps — sparse signal
2026-02 2602.12229 VMPO Direct Preference SMC/variance perspective on diffusion alignment unexplored
2026-03 2603.07700 TDM-R1 Direct Preference RL for few-step (2–4 step) diffusion with non-differentiable rewards
2026-03 2603.21872 SAGE-GRPO Policy Gradient SDE noise in video generation degrades rollout quality off-manifold

Policy Gradient Paradigm Advances

Papers whose objective is a PPO-clip / importance-weighted policy gradient over the denoising trajectory (the SDE-based GRPO family) — or, more broadly, value/Q-based and reward-maximisation variants that optimise the same expected-reward objective.


BranchGRPO — 2509.06040 · Sep 2025

GitHub: https://github.com/Fredreic1849/BranchGRPO

Problem: Sequential rollouts in GRPO are slow; applying the same trajectory-level reward to all timesteps gives no per-step credit signal; low-value branches waste compute.

Approach: Structures denoising as a branching tree rooted at a shared prefix. Branches split at selected timesteps to produce multiple candidate continuations from one shared trunk, amortizing early-step compute. Rewards are propagated upward via path-probability weighting to produce per-depth advantages:

\[\hat A_t^{(i)} = \frac{\sum_{b: t \in b} p(b)R^{(b)}}{\sum_{b: t \in b} p(b)} - \overline R\]

Only nodes in high-value branches receive gradient. Achieves ~55% compute reduction and +16% HPDv2.1 alignment vs. FlowGRPO.


TaRoS — Target-Robust Reward Signaling · 2511.19356 · Nov 2025

Problem: Under sustained GRPO training, reward models over-fit to shortcut patterns (Goodhart's Law), especially in video generation where rewards span multiple quality axes. Reward saturation within groups removes the advantage signal.

Approach: Intra-group sparse filtering across reward components: decomposes the scalar reward into \(K\) aspect scores \(\lbrace r_k^{(i)}\rbrace\) and retains per-aspect advantage only for the top-\(m\) samples per group (sparsity mask \(\mathcal{S}_k\)):

\[\hat A_k^{(i)} = \mathbf{1}[i \in \mathcal{S}_k]\cdot \frac{r_k^{(i)} - \overline r_k}{\text{std}(\lbrace r_k^{(j)}\rbrace) + \delta}\]

Combined advantage: \(\hat A^{(i)} = \sum_k w_k \hat A_k^{(i)}\). Prevents any single reward axis from saturating; targets the weakest aspects per group.


TreeGRPO — 2512.08153 · Dec 2025

Website: https://treegrpo.github.io

Problem: GRPO reuses the same scalar advantage for every denoising step in a trajectory — no fine-grained credit. BranchGRPO reduces compute but still uses path-aggregated rewards rather than true step-level estimates.

Approach: Recasts the full denoising process as a search tree. A single root noise \(x_T\) branches into multiple children at strategically chosen timesteps; rewards are back-propagated through the tree to assign node-specific advantages:

\[\hat A_t^\text{node} = R(\text{subtree}(t)) - V(x_t)\]

where \(V(x_t)\) is estimated by averaging sibling rewards. Multi-child branching yields multiple policy updates per forward pass. Achieves 2.4× training speedup vs. GRPO.


Pro-GRPO — Expand and Prune · 2512.15347 · Dec 2025

Problem: Large group sizes \(N\) reduce advantage variance but scale linearly in compute. Many samples cluster near the group mean (reward clustering), providing zero signal while consuming resources.

Approach: Optimal Variance Filtering (OVF): first expand to a large candidate group \(N_\text{large}\), then prune mid-generation (not post-hoc) using latent features to identify and discard reward-clustered trajectories. The surviving group \(\lbrace x_0^{(i)}\rbrace_{i \in \mathcal{S}}\) has much higher per-sample variance:

\[\mathcal{S} = \arg\max_{|\mathcal{S}|=N_\text{keep}} \text{Var}(\lbrace r^{(i)}\rbrace_{i \in \mathcal{S}})\]

OVF uses intermediate latent features (not final images) to predict reward clustering, enabling early abortion of low-signal branches.


SuperFlow — 2512.17951 · Dec 2025

Problem: FlowGRPO uses the same group size \(N\) per prompt regardless of per-prompt reward variance. Trajectory-level advantages are biased because they ignore continuous-time flow dynamics at the step level.

Key contributions: 1. Dynamic-Group sampling: group size \(N_c\) adapts per prompt based on estimated reward standard deviation \(\hat\sigma_c\):

$\(N_c = \left\lceil N_0 \cdot \frac{\hat\sigma_c}{\overline\sigma}\right\rceil\)$ 2. Step-level advantage: derives per-step advantage estimates from the flow ODE dynamics (continuous-time stochastic calculus) rather than reusing the terminal advantage at all steps.

Achieves +38.4% GenEval improvement over SD3.5-M baseline.


DiverseGRPO — 2512.21514 · Dec 2025

Website: https://henglin-liu.github.io/DiverseGRPO/

Problem: Standard GRPO collapses to a few high-reward visual modes; the generated distribution loses semantic and stylistic diversity even as proxy reward improves.

Key contributions: 1. Cluster-based exploratory reward: spectral clustering over per-prompt generated images; samples in under-represented clusters receive a diversity bonus:

$\(r_\text{div}^{(i)} = \frac{1}{|\mathcal{C}(i)|} \cdot r_\text{quality}^{(i)}\)$ where \(\mathcal{C}(i)\) is the cluster containing image \(i\). Smaller clusters → larger bonus. 2. Structure-aware regularisation: stronger KL penalty at early denoising steps (high-\(t\)) to preserve structural diversity; relaxed at late steps where fine-grained reward signal matters.


E-GRPO — Entropy-Aware GRPO · 2601.00423 · Jan 2026

GitHub: https://github.com/shengjun-zhang/VisualGRPO

Problem: GRPO applies SDE noise uniformly to all denoising steps. Low-entropy steps (where the model is near-deterministic) add noise but produce nearly indistinguishable rollouts, diluting reward signal. High-entropy steps (where the model is most uncertain) are under-exploited.

Approach: Merge consecutive low-entropy steps into one effective high-entropy SDE step; apply ODE to the remainder. Entropy is measured via the velocity field divergence \(\nabla \cdot v_\theta\):

\[H_t = -\mathbb{E}[\log \pi_\theta(x_{t-\Delta t} | x_t)] \approx \frac{d}{2}\log(2\pi e\sigma_t^2\Delta t)\]

Steps with \(H_t < H_\text{thresh}\) are merged. The merged step's noise is scaled to preserve total stochasticity:

\[\sigma_\text{merged} = \sqrt{\sum_{k \in \text{merged}} \sigma_{t_k}^2\Delta t}\]

Focuses exploration budget where it has most discriminative effect on reward.


TAGRPO — Trajectory-Aligned GRPO for I2V · 2601.05729 · Jan 2026

Problem: Standard GRPO for image-to-video (I2V) generation provides only trajectory-level scalar advantages; video quality depends on temporal alignment between frames, which scalar rewards poorly capture.

Approach: Contrastive trajectory alignment in latent space. Within each group, high-advantage video latent trajectories are attracted toward each other; low-advantage trajectories are repelled from high-advantage ones:

\[\mathcal{L}_\text{contrast}(\theta) = -\sum_{i \in \mathcal{G}^{+}}\sum_{j \in \mathcal{G}^-} \log \frac{\exp(d(\mathbf{z}^{(i)}, \mathbf{z}^{(j)}) / \tau)}{\sum_{k} \exp(d(\mathbf{z}^{(i)}, \mathbf{z}^{(k)}) / \tau)}\]

where \(\mathbf{z}^{(i)} = \lbrace x_{t_k}^{(i)}\rbrace_{t_k \in T_\text{SDE}}\) is the latent trajectory and \(d(\cdot,\cdot)\) is a trajectory similarity. Combined with GRPO loss and a memory bank for diversity.


DRIFT — Beyond Dirac Delta · 2601.12401 · Jan 2026

Problem: RL fine-tuning of diffusion models is provably converging to a Dirac delta (mode collapse) under standard PPO/GRPO objectives — demonstrated theoretically and empirically. No existing method has a principled fix.

Approach: Three-pronged diversity incentivisation within the GRPO framework: 1. Reward-concentrated sampling: retain only trajectories with reward \(r^{(i)} \in [\overline r - \alpha\sigma, \overline r + \alpha\sigma]\) for gradient computation (removes outlier-induced mode pull). 2. Stochastic prompt variation: augment the conditioning \(c\) with random variation \(\tilde c = c + \delta c\) during rollout to expand the conditioning manifold. 3. Potential-based reward shaping: add a diversity potential \(\Phi(x_0^{(i)})\) to the reward that measures distance to previously-generated samples, preventing the policy from collapsing to already-visited modes:

\[\tilde r^{(i)} = r^{(i)} + \lambda_\Phi\Phi(x_0^{(i)}), \quad \Phi(x_0) = \min_{j \in \mathcal{M}} d(x_0, x_0^{(j)})\]

DenseGRPO — 2601.20218 · Jan 2026

Problem: All GRPO variants assign the same terminal reward \(R(x_0, c)\) to every denoising step — a sparse credit assignment problem. The model cannot learn which steps contributed most to the final quality.

Key contributions: 1. ODE-based intermediate reward estimation: at each step \(t_k\), compute the Tweedie clean-image prediction \(\hat x_0(x_{t_k})\) and evaluate the reward model on it as a proxy for step \(t_k\)'s contribution:

$\(r_{t_k}^{(i)} = R\left(\hat x_0(x_{t_k}^{(i)}), c\right), \quad \hat x_0 = x_t - tv_\theta(x_t,t,c)\)$ 2. Reward-aware stochasticity calibration: set the SDE noise coefficient \(\sigma_{t_k}\) proportional to the expected reward gradient magnitude at step \(t_k\), focusing exploration at high-impact steps.

Step advantage:

\[\hat A_{t_k}^{(i)} = \frac{r_{t_k}^{(i)} - \overline r_{t_k}}{\text{std}(\lbrace r_{t_k}^{(j)}\rbrace) + \delta}\]

SAGE-GRPO — 2603.21872 · Mar 2026

GitHub: https://github.com/Tencent-Hunyuan/SAGE-GRPO

Problem: For video generation, ODE-to-SDE conversion injects noise that pushes intermediate states off the video data manifold; reward models trained on clean videos cannot reliably score these artifact-laden samples, destabilising RL training.

Approach: Defines the pre-trained flow model as implicitly specifying a video manifold and enforces dual proximity constraints: - Micro-constraint (per step): the SDE noise is clipped so \(x_{t-\Delta t}^\text{SDE}\) stays within \(\delta_\text{micro}\) of the ODE trajectory \(x_{t-\Delta t}^\text{ODE}\):

$\(\Vert x_{t-\Delta t}^\text{SDE} - x_{t-\Delta t}^\text{ODE}\Vert_2 \leq \delta_\text{micro}\)$ - Macro-constraint (rollout level): terminal frames must pass a quality gate (\(R(x_0) \geq \tau_\text{macro}\)) before contributing to the gradient.

Combined with FlowGRPO-style GRPO objective over the SDE window.


Direct Preference Paradigm Advances

Papers whose objective is a preference / ELBO / matching loss on final (or single-step) samples — solver-agnostic, no per-step importance ratio. A few entries here are actor-critic, soft-Q, or reward-maximisation variants that are solver-agnostic but reward-based rather than preference-based per se; they are grouped here pending dedicated review and flagged in their entries.


DAV — Diffusion Alignment as Variational EM · 2510.00502 · Oct 2025

GitHub: https://github.com/Jaewoopudding/dav

Problem: Standard KL-regularised RL for diffusion models converges to a single mode of the reward-tilted distribution (mode-seeking reverse KL). Multi-modal reward landscapes require a mode-covering objective.

Approach: Frames diffusion alignment as iterative EM: - E-step (test-time search): find a set of diverse high-reward samples from the variational posterior \(q(x_0) \propto r(x_0) \cdot \pi_{\theta_\text{old}}(x_0)\) via MCMC or rejection sampling. - M-step (amortisation): minimise forward KL (mode-covering) from \(q\) into \(\pi_\theta\):

\[\mathcal{L}_\text{M}(\theta) = D_\text{KL}(q \Vert \pi_\theta) = -\mathbb{E}_{x_0 \sim q}[\log \pi_\theta(x_0)] + \text{const}\]

The forward KL is approximated via the ELBO (same trick as Diffusion-DPO):

\[\mathcal{L}_\text{M}(\theta) \approx \mathbb{E}_{x_0 \sim q,t,\epsilon}\left[\Vert v_\theta(x_t,t,c) - u_t\Vert^2\right]\]

Works for both continuous (T2I) and discrete (DNA) domains without differentiable reward.


P-GRAFT — Intermediate Distribution Shaping · 2510.02692 · Oct 2025

Venue: ICLR 2026

Problem: RAFT-based methods (rejection sampling fine-tuning) shape only the terminal distribution \(\pi(x_0)\); shaping at intermediate noise levels is theoretically suboptimal; no unified theory connects RAFT variants.

Approach: Introduces the GRAFT framework (Generalised RAFT) showing that all RAFT variants implicitly minimise:

\[\mathcal{L}_\text{GRAFT}(\theta) = D_\text{KL}(\tilde\pi \Vert \pi_\theta), \quad \tilde\pi(x_0) \propto \exp(R(x_0)/\beta)\pi_\text{ref}(x_0)\]

P-GRAFT extends this by shaping intermediate distributions \(\tilde\pi_t(x_t)\) at noise level \(t\), not just at \(t=0\):

\[\mathcal{L}_\text{P-GRAFT}(\theta) = \sum_t \lambda_tD_\text{KL}(\tilde\pi_t \Vert \pi_\theta^t)\]

The bias-variance tradeoff is analysed: intermediate shaping has higher bias but lower variance than terminal-only shaping. Extended to flow model error correction.


AC-Flow — Actor-Critic for Flow Matching · 2510.18072 · Oct 2025

Website: https://www.jiajunfan.com/projects/ac-flow/

Problem: Flow matching models lack a principled actor-critic framework that provides intermediate (per-step) feedback; naive critic regression on cumulative rewards is unstable.

Approach: - Critic: trains a value function \(V_\phi(x_t, t, c)\) on intermediate states via reward shaping for stable convergence:

$\(V_\phi(x_t,t,c) = \mathbb{E}_{x_0 | x_t}\left[R(x_0,c)\right]\)$ - Wasserstein diversity regularisation: penalises the 2-Wasserstein distance between the generated distribution and a reference in reward-weighted feature space:

$\(\mathcal{L}_\text{div}(\theta) = W_2(\pi_\theta,\pi_\text{ref})\)$ - Dual stability: advantage clipping + critic warm-up to prevent Q-value collapse in early training.

Actor update uses the intermediate advantage \(\hat A_t^{(i)} = R^{(i)} - V_\phi(x_t^{(i)}, t, c)\) weighted by the flow matching loss.


DDRL — Data-Regularized Diffusion RL · 2512.04332 · Dec 2025

GitHub: https://github.com/nvidia-cosmos/cosmos-rl

Problem: On-policy KL regularisation against a fixed reference policy is the primary cause of reward hacking (quality degradation, over-stylisation, diversity loss) in DDPO/FlowGRPO-style training. This is confirmed at scale (>1M GPU-hours of video generation training).

Approach: Replace on-policy KL with forward KL to an off-policy data distribution \(p_\text{data}\). The standard diffusion training loss is exactly the forward KL:

\[\mathcal{L}_\text{diff}(\theta) = \mathbb{E}_{x_0 \sim p_\text{data},t,\epsilon}\left[\Vert v_\theta(x_t,t,c) - u_t\Vert^2\right] = D_\text{KL}(p_\text{data} \Vert \pi_\theta) + \text{const}\]

Combined objective:

\[\mathcal{L}_\text{DDRL}(\theta) = -\mathbb{E}_{x_0 \sim \pi_\theta}[R(x_0,c)] + \alpha\mathcal{L}_\text{diff}(\theta)\]

The reward term is on-policy; the data-regularisation term is off-policy, providing an unbiased anchor that prevents reward hacking. Scales to image and video generation.


SQDF — Soft Q-based Diffusion Finetuning · 2512.04559 · Dec 2025

GitHub: https://github.com/Shin-woocheol/SQDF

Problem: Diffusion RL methods lack a principled soft Q-function for credit assignment over the long denoising chain; temporal discounting and multi-step value estimation are absent.

Key contributions: 1. Training-free soft Q-function: estimate \(Q_\text{soft}(x_t, a_t)\) from the current model via the Bellman equation, using consistency models to efficiently predict \(x_0\) from \(x_t\) for value bootstrapping. 2. Discounted denoising: introduces discount factor \(\gamma < 1\) so later denoising steps (closer to \(x_0\)) receive higher weight:

$\(Q_t = r_t + \gamma V_{t-1}, \quad V_t = \mathbb{E}_{x_{t-1}|x_t}[Q_{t-1}]\)$ 3. Off-policy replay buffer: stores (trajectory, reward) pairs across iterations; combined with on-policy samples to improve mode coverage.

Policy gradient: \(\nabla_\theta \mathcal{L} = \mathbb{E}[\nabla_\theta \log \pi_\theta(x_{t-1}|x_t) \cdot Q_\text{soft}(x_t, x_{t-1})]\).


GARDO — Gated Adaptive Regularization · 2512.24138 · Dec 2025

GitHub: https://github.com/tinnerhrhe/GARDO

Problem: KL regularisation prevents reward hacking but also impedes exploration — a fundamental tension. Applying the same penalty everywhere regardless of local policy uncertainty is suboptimal.

Approach: 1. Uncertainty-gated KL: measure per-sample uncertainty \(u^{(i)} = \text{Var}_{t,\epsilon}[\hat A^{(i)}(x_t)]\); apply KL penalty only where \(u^{(i)} > u_\text{thresh}\):

$\(\mathcal{L}_\text{GARDO}(\theta) = \mathcal{L}_\text{GRPO}(\theta) - \beta\mathbb{E}\left[\mathbf{1}[u^{(i)} > u_\text{thresh}]D_\text{KL}(\pi_\theta(x_0)\Vert\pi_\text{ref}(x_0))\right]\)$ 2. Diversity-aware advantage shaping: adds a novelty bonus to \(\hat A^{(i)}\) for samples that differ from recently-generated images, encouraging exploration of new modes.

Acts as a plug-in wrapper compatible with any base GRPO/ELBO algorithm.


VMPO — Variance Minimisation Policy Optimisation · 2602.12229 · Feb 2026

Problem: The standard KL-based alignment objective has been thoroughly studied, but the variance of the importance weights (a key driver of instability in SMC-type methods) has not been used as a policy optimisation target.

Approach: Frames diffusion alignment as Sequential Monte Carlo (SMC): the denoising model is a proposal \(\pi_\theta\); the reward-tilted distribution \(\tilde\pi \propto R \cdot \pi_\text{ref}\) is the target; importance weights \(w^{(i)} = \tilde\pi(x_0^{(i)})/\pi_\theta(x_0^{(i)})\) measure how well \(\pi_\theta\) covers \(\tilde\pi\).

VMPO objective: minimise the variance of log importance weights:

\[\mathcal{L}_\text{VMPO}(\theta) = \text{Var}_{x_0 \sim \pi_\theta}\left[\log\frac{\tilde\pi(x_0)}{\pi_\theta(x_0)}\right]\]

Key theorem: under on-policy sampling, \(\nabla_\theta \mathcal{L}_\text{VMPO} = \nabla_\theta D_\text{KL}(\tilde\pi \Vert \pi_\theta)\), establishing equivalence with KL alignment on-policy. Off-policy, VMPO provides a correction term that reduces variance without mode-seeking bias.


TDM-R1 — Trajectory Distribution Matching · 2603.07700 · Mar 2026

GitHub: https://github.com/Luo-Yihong/TDM-R1

Problem: RL for few-step (2–4 step) diffusion models. Non-differentiable rewards (binary human preference, object counts, OCR correctness) cannot be directly backpropagated; no surrogate reward learning framework exists for fast ODE samplers.

Approach: Decouples into two stages: 1. Surrogate reward learning via DGPO-style group preference optimisation on deterministic ODE trajectories. For each prompt, generate a group of \(N\) images \(\lbrace x_0^{(i)}\rbrace\); the surrogate reward \(\hat R_\psi\) is trained to predict group-relative rankings using the ELBO log-ratio as the policy model:

$\(\mathcal{L}_\text{surr}(\psi) = -\mathbb{E}\left[\log \sigma\left(\hat R_\psi(x_0^{+}) - \hat R_\psi(x_0^-)\right)\right]\)$ 2. Generator training guided by the surrogate:

\[\mathcal{L}_\text{gen}(\theta) = -\mathbb{E}_{x_0 \sim \pi_\theta}\left[\hat R_\psi(x_0,c)\right] + \beta D_\text{KL}(\pi_\theta \Vert \pi_\text{ref})\]

The key insight: ODE trajectory intermediate states \(\lbrace x_{t_k}\rbrace\) provide unbiased, artifact-free intermediate value estimates (no SDE needed). Achieves GenEval 61% → 92% with only 4 denoising steps, surpassing the 40-step base model (63%) and GPT-4o (84%).


Cross-Cutting Notes

The exploration–exploitation tension is now the central open problem. DiverseGRPO, DRIFT, GARDO, and DenseGRPO all attack it from different angles (diversity shaping, potential-based rewards, uncertainty-gated KL, dense credit assignment). No clear winner yet.

Dense credit assignment (TaRoS, DenseGRPO, TreeGRPO) is an active area — the mismatch between where gradients flow (denoising steps) and where reward is observed (terminal image) remains unsolved at the fundamental level.

Few-step / distilled models (TDM-R1) open a new front: RL methods optimized for many-step samplers may not transfer to 2–4 step consistency/distilled models.

Video generation RL (TaRoS, TAGRPO, SAGE-GRPO) is becoming a distinct sub-field as temporal consistency and off-manifold noise artifacts introduce qualitatively new challenges vs. image generation.