Iso-GRPO: Temporal & Normalization Methods

• Iso-GRPO is a family of GRPO variants that impose iso constraints—whether temporal, reward-scale, or disagreement normalization—to ensure fair comparisons within grouped samples.
• In video diffusion, the Flash-GRPO method uses iso-temporal grouping to synchronize rollouts at a single timestep, reducing variance and boosting training stability with Temporal Gradient Rectification.
• Iso-GRPO techniques extend to various domains by leveraging invariance under affine reward transformations, enabling robust, critic-free policy optimization with efficient computation.

Iso-GRPO is a label used in recent arXiv literature for GRPO variants or interpretations that impose an “iso” constraint on how group-relative comparisons are formed or understood. In its most explicit published usage, it denotes the iso-temporal component of Flash-GRPO for video diffusion models: for each prompt, all rollouts within a GRPO group are forced to share the same diffusion timestep, and the resulting one-step update is combined with Temporal Gradient Rectification (He et al., 15 May 2026). Other papers use the term more loosely to refer to the affine reward invariance induced by GRPO’s shift-and-scale normalization, or to hypothetical variants that preserve such invariances while modifying the KL or group construction (Sepúlveda et al., 9 Jun 2026, Vojnovic et al., 25 Feb 2025). This suggests that Iso-GRPO is best treated as a family resemblance concept rather than a single universally standardized algorithm.

1. Terminological scope and emergence

The clearest algorithmic instantiation of Iso-GRPO appears in “Flash-GRPO: Efficient Alignment for Video Diffusion via One-Step Policy Optimization,” where “Iso-GRPO” is essentially Flash-GRPO’s iso-temporal GRPO: for each prompt, every rollout in the group uses the same diffusion timestep, so the within-group comparison is not confounded by timestep difficulty (He et al., 15 May 2026). In that setting, “iso” refers to temporal consistency inside the group.

A second usage arises in analyses of GRPO’s normalization itself. In “Baseline-Free Policy Optimization for Neural Combinatorial Optimization,” the update induced by group-relative normalization is described as “invariant to affine reward scaling within each instance,” and that invariance is explicitly linked to the kind of “iso-/scale invariance” associated with labels such as “Iso-GRPO” in the alignment literature (Sepúlveda et al., 9 Jun 2026). “What is the Alignment Objective of GRPO?” makes the same point more formally: the group-relative advantage produced by subtracting the group mean and dividing by the group standard deviation is invariant under any affine reward transformation $r_i \mapsto a r_i + b$ with $a>0$ (Vojnovic et al., 25 Feb 2025).

The published record therefore supports two complementary senses of the term. One is operational and refers to iso-temporal grouping in video diffusion. The other is structural and refers to invariance properties of group-normalized GRPO objectives. The coexistence of these meanings is important, because many claims about Iso-GRPO in later work are really claims about one of these two aspects rather than about a single canonical algorithm.

2. GRPO as the substrate

Any account of Iso-GRPO begins with GRPO itself. In the critic-free formulation used for robust blind interference alignment with fluid antennas, GRPO optimizes policy parameters $w$ via a PPO-style clipped surrogate with a KL regularizer to a fixed reference policy $\pi_{\mathrm{ref}}$:

$$\begin{aligned} J_\text{GRPO}(w) = \mathbb{E}\Bigg[\frac{1}{G} \sum_{g=1}^G \frac{1}{O_g} \sum_{t=1}^{O_g} \Big\{ &\min \Big( \frac{\pi_w(o_{g,t}|a,o_{g,<t})}{\pi_{w_\text{old}}(o_{g,t}|a,o_{g,<t})}, \ &\text{clip}\Big(\frac{\pi_w(o_{g,t}|a,o_{g,<t})}{\pi_{w_\text{old}}(o_{g,t}|a,o_{g,<t})},1-c,1+c\Big)\Big)\,\widehat{A}_{g} \ &\qquad - \mu \,\mathbb{D}_\text{KL}[\pi_{w}\|\pi_\text{ref}] \Big\}\Bigg]. \end{aligned}$$

Its central innovation is the group-relative advantage

$$\widehat{A}_{g} = \frac{r_g - \text{mean}(\{r_1,\ldots,r_G\})}{\text{std}(\{r_1,\ldots,r_G\})},$$

where $r_g$ is a trajectory-wise reward and the same scalar $\widehat{A}_g$ is reused for all timesteps in trajectory $g$ (Peng et al., 20 Jan 2026).

This differs from standard PPO in three ways. First, there is no critic network and no value-function estimation. Second, the learning signal is trajectory-wise and group-relative rather than per-timestep and critic-based. Third, stability is provided by both PPO-style clipping relative to $\pi_{w_{\text{old}}}$ and an explicit KL penalty to a separate reference policy $a>0$0 (Peng et al., 20 Jan 2026).

The practical consequence is a critic-free actor-only architecture. In the fluid-antenna application, PPO uses actor and critic networks with identical MLP architecture, whereas GRPO uses only the actor. That paper reports a model-size reduction of 49.6%, a FLOP reduction of 46.7%, performance gains of 4.17% over PPO and 30.29% over a 100K-step PPO-Init baseline, and substantially larger gains over heuristic MaximumGain and RandomGain baselines (Peng et al., 20 Jan 2026). Although that work does not use the label Iso-GRPO, it provides a clean reference point for the algorithmic core from which later “iso” variants depart.

3. Iso-temporal grouping in video diffusion

In video diffusion, the main obstacle to naïve one-step GRPO is timestep-confounded variance. If each rollout in a group is updated at its own sampled timestep $a>0$1, then reward differences within the group mix two effects: policy quality and intrinsic timestep difficulty. Flash-GRPO resolves this with iso-temporal grouping. For each prompt $a>0$2, a single timestep is sampled,

$a>0$3

and the group is constructed as

$a>0$4

so that all group members share the same diffusion timestep and differ only through different Gaussian noise samples (He et al., 15 May 2026).

The paper’s variance argument is explicit. With naïve one-step sampling, the reward variance contains a term

$a>0$5

which is precisely the variance induced by timestep difficulty. Iso-temporal grouping removes this as a source of within-group variation because $a>0$6 is constant inside each group. The groupwise baseline is then a fair estimate of typical performance at that prompt and that timestep, rather than an average over heterogeneous denoising difficulties (He et al., 15 May 2026).

Flash-GRPO couples this construction with Temporal Gradient Rectification. In the reverse-SDE update, the policy gradient contains a time-dependent scaling factor $a>0$7, so gradients at different timesteps can differ by orders of magnitude for reasons unrelated to learning value. The paper therefore defines an unclipped rectified loss

$a>0$8

thereby dividing out the hidden time-dependent multiplier (He et al., 15 May 2026). Iso-temporal grouping addresses the comparability of samples inside a group; TGR addresses the comparability of gradients across groups assigned to different timesteps.

Operationally, Flash-GRPO is a one-step method. For each prompt group, only the single timestep $a>0$9 is stochastic and receives gradients; all other timesteps are deterministic ODE steps. The paper describes the resulting objective as a Monte Carlo estimator of the full-trajectory sum over timesteps, while emphasizing that the video-level reward is tied back to the single stochastic reverse step (He et al., 15 May 2026).

Empirically, the method is presented as both more stable and more efficient than sliding-window baselines and full-trajectory GRPO. On 1.3B to 14B parameter models it is reported to deliver substantial training acceleration with consistent stability and state-of-the-art alignment quality. In the 1.3B ablation, naïve single-step training attains HPSv3 reward 4.64, adding iso-temporal grouping raises this to 5.31, and combining iso-temporal grouping with TGR yields 5.42 together with improved training stability. On Wan2.1-T2V-1.3B using 350 GPU-hours, Flash-GRPO records Aesthetic Quality 66.43, Subject consistency 98.70, and Object class 90.00; the paper also reports around $w$0 acceleration in training cost relative to full-trajectory training (He et al., 15 May 2026).

4. Invariance, alignment objective, and the “iso” interpretation

A distinct line of work interprets “Iso-GRPO” through invariance rather than through timestep control. In GRPO’s canonical reward normalization, each sampled output $w$1 receives the advantage

$w$2

Because both the mean and standard deviation transform covariantly, $w$3 is invariant under affine reward transformations $w$4 with $w$5 (Vojnovic et al., 25 Feb 2025). The NCO study makes the same point operationally: within each instance, if rewards are rescaled as $w$6, the normalized group-relative advantage remains approximately unchanged up to the $w$7 floor, so the update depends on relative ranking rather than absolute reward scale (Sepúlveda et al., 9 Jun 2026).

Theoretical analysis of GRPO’s stationary points shows that this invariance is coupled to a nonstandard preference aggregation rule. Ignoring clipping, the GRPO reward term can be written as an expectation over a group-relative preference function $w$8, and the stationary policy satisfies

$w$9

This produces a rational rescaling of the reference policy rather than the exponential reweighting familiar from direct-KL RLHF (Vojnovic et al., 25 Feb 2025). The same paper shows that the KL surrogate used in GRPO behaves, at stationarity, as reverse KL $\pi_{\mathrm{ref}}$0, not as direct KL $\pi_{\mathrm{ref}}$1. In that sense, the invariance of the reward model and the mode-seeking character of the penalty are jointly constitutive of the GRPO objective.

Special cases sharpen the connection to “iso” interpretations. For $\pi_{\mathrm{ref}}$2, the normalized advantage collapses to a sign comparison, so the reward model reduces to pairwise comparison preferences. In the large-$\pi_{\mathrm{ref}}$3 limit, the reward term becomes a z-scored expected reward under the old policy (Vojnovic et al., 25 Feb 2025). The paper also analyzes direct-KL and shift-only variants, showing that direct KL restores logarithmic pooling, while removing scale normalization breaks scale invariance.

The NCO experiments provide concrete evidence that these invariances can matter algorithmically. In a controlled comparison on TSP and CVRP, GRPO is presented as a baseline-free method with advantages computed purely from within-group z-scores. At matched gradient updates, it achieves solution quality within 2% of POMO, avoids the TSP-100 training collapse observed with REINFORCE, and is explicitly described as “invariant to affine reward scaling within each instance” (Sepúlveda et al., 9 Jun 2026). This suggests that one strand of Iso-GRPO discourse is best read as shorthand for GRPO variants that preserve reward-scale or preference-order invariance while changing other parts of the objective.

5. Analytical reinterpretations: disagreement weighting and hidden process rewards

Recent theory has supplied two additional ways to parse what an Iso-GRPO variant would actually control. The first is the group-standard-deviation identity for binary rewards. For a prompt with $\pi_{\mathrm{ref}}$4 sampled answers, $\pi_{\mathrm{ref}}$5 correct responses, mean reward $\pi_{\mathrm{ref}}$6, and standard deviation

$\pi_{\mathrm{ref}}$7

the GRPO per-prompt update

$\pi_{\mathrm{ref}}$8

can be written exactly as

$\pi_{\mathrm{ref}}$9

where $$\begin{aligned} J_\text{GRPO}(w) = \mathbb{E}\Bigg[\frac{1}{G} \sum_{g=1}^G \frac{1}{O_g} \sum_{t=1}^{O_g} \Big\{ &\min \Big( \frac{\pi_w(o_{g,t}|a,o_{g,<t})}{\pi_{w_\text{old}}(o_{g,t}|a,o_{g,<t})}, \ &\text{clip}\Big(\frac{\pi_w(o_{g,t}|a,o_{g,<t})}{\pi_{w_\text{old}}(o_{g,t}|a,o_{g,<t})},1-c,1+c\Big)\Big)\,\widehat{A}_{g} \ &\qquad - \mu \,\mathbb{D}_\text{KL}[\pi_{w}\|\pi_\text{ref}] \Big\}\Bigg]. \end{aligned}$$0 and $$\begin{aligned} J_\text{GRPO}(w) = \mathbb{E}\Bigg[\frac{1}{G} \sum_{g=1}^G \frac{1}{O_g} \sum_{t=1}^{O_g} \Big\{ &\min \Big( \frac{\pi_w(o_{g,t}|a,o_{g,<t})}{\pi_{w_\text{old}}(o_{g,t}|a,o_{g,<t})}, \ &\text{clip}\Big(\frac{\pi_w(o_{g,t}|a,o_{g,<t})}{\pi_{w_\text{old}}(o_{g,t}|a,o_{g,<t})},1-c,1+c\Big)\Big)\,\widehat{A}_{g} \ &\qquad - \mu \,\mathbb{D}_\text{KL}[\pi_{w}\|\pi_\text{ref}] \Big\}\Bigg]. \end{aligned}$$1 are the mean score vectors for correct and incorrect responses. Dr. GRPO replaces this with $$\begin{aligned} J_\text{GRPO}(w) = \mathbb{E}\Bigg[\frac{1}{G} \sum_{g=1}^G \frac{1}{O_g} \sum_{t=1}^{O_g} \Big\{ &\min \Big( \frac{\pi_w(o_{g,t}|a,o_{g,<t})}{\pi_{w_\text{old}}(o_{g,t}|a,o_{g,<t})}, \ &\text{clip}\Big(\frac{\pi_w(o_{g,t}|a,o_{g,<t})}{\pi_{w_\text{old}}(o_{g,t}|a,o_{g,<t})},1-c,1+c\Big)\Big)\,\widehat{A}_{g} \ &\qquad - \mu \,\mathbb{D}_\text{KL}[\pi_{w}\|\pi_\text{ref}] \Big\}\Bigg]. \end{aligned}$$2, and DAPO discards the $$\begin{aligned} J_\text{GRPO}(w) = \mathbb{E}\Bigg[\frac{1}{G} \sum_{g=1}^G \frac{1}{O_g} \sum_{t=1}^{O_g} \Big\{ &\min \Big( \frac{\pi_w(o_{g,t}|a,o_{g,<t})}{\pi_{w_\text{old}}(o_{g,t}|a,o_{g,<t})}, \ &\text{clip}\Big(\frac{\pi_w(o_{g,t}|a,o_{g,<t})}{\pi_{w_\text{old}}(o_{g,t}|a,o_{g,<t})},1-c,1+c\Big)\Big)\,\widehat{A}_{g} \ &\qquad - \mu \,\mathbb{D}_\text{KL}[\pi_{w}\|\pi_\text{ref}] \Big\}\Bigg]. \end{aligned}$$3 groups entirely (Bay et al., 30 Jun 2026). The paper’s main claim is that GRPO, Dr. GRPO, and DAPO are “three operations on one number,” namely the group standard deviation. This suggests that any Iso-GRPO-like attempt to make updates more isotropic across prompts can be read as choosing a different effective function $$\begin{aligned} J_\text{GRPO}(w) = \mathbb{E}\Bigg[\frac{1}{G} \sum_{g=1}^G \frac{1}{O_g} \sum_{t=1}^{O_g} \Big\{ &\min \Big( \frac{\pi_w(o_{g,t}|a,o_{g,<t})}{\pi_{w_\text{old}}(o_{g,t}|a,o_{g,<t})}, \ &\text{clip}\Big(\frac{\pi_w(o_{g,t}|a,o_{g,<t})}{\pi_{w_\text{old}}(o_{g,t}|a,o_{g,<t})},1-c,1+c\Big)\Big)\,\widehat{A}_{g} \ &\qquad - \mu \,\mathbb{D}_\text{KL}[\pi_{w}\|\pi_\text{ref}] \Big\}\Bigg]. \end{aligned}$$4 on disagreement magnitude.

The same paper connects disagreement weighting to problem difficulty. In the large-group limit, GRPO’s expected gradient magnitude scales like $$\begin{aligned} J_\text{GRPO}(w) = \mathbb{E}\Bigg[\frac{1}{G} \sum_{g=1}^G \frac{1}{O_g} \sum_{t=1}^{O_g} \Big\{ &\min \Big( \frac{\pi_w(o_{g,t}|a,o_{g,<t})}{\pi_{w_\text{old}}(o_{g,t}|a,o_{g,<t})}, \ &\text{clip}\Big(\frac{\pi_w(o_{g,t}|a,o_{g,<t})}{\pi_{w_\text{old}}(o_{g,t}|a,o_{g,<t})},1-c,1+c\Big)\Big)\,\widehat{A}_{g} \ &\qquad - \mu \,\mathbb{D}_\text{KL}[\pi_{w}\|\pi_\text{ref}] \Big\}\Bigg]. \end{aligned}$$5 for a prompt with success probability $$\begin{aligned} J_\text{GRPO}(w) = \mathbb{E}\Bigg[\frac{1}{G} \sum_{g=1}^G \frac{1}{O_g} \sum_{t=1}^{O_g} \Big\{ &\min \Big( \frac{\pi_w(o_{g,t}|a,o_{g,<t})}{\pi_{w_\text{old}}(o_{g,t}|a,o_{g,<t})}, \ &\text{clip}\Big(\frac{\pi_w(o_{g,t}|a,o_{g,<t})}{\pi_{w_\text{old}}(o_{g,t}|a,o_{g,<t})},1-c,1+c\Big)\Big)\,\widehat{A}_{g} \ &\qquad - \mu \,\mathbb{D}_\text{KL}[\pi_{w}\|\pi_\text{ref}] \Big\}\Bigg]. \end{aligned}$$6, whereas Dr. GRPO optimizes the raw success rate $$\begin{aligned} J_\text{GRPO}(w) = \mathbb{E}\Bigg[\frac{1}{G} \sum_{g=1}^G \frac{1}{O_g} \sum_{t=1}^{O_g} \Big\{ &\min \Big( \frac{\pi_w(o_{g,t}|a,o_{g,<t})}{\pi_{w_\text{old}}(o_{g,t}|a,o_{g,<t})}, \ &\text{clip}\Big(\frac{\pi_w(o_{g,t}|a,o_{g,<t})}{\pi_{w_\text{old}}(o_{g,t}|a,o_{g,<t})},1-c,1+c\Big)\Big)\,\widehat{A}_{g} \ &\qquad - \mu \,\mathbb{D}_\text{KL}[\pi_{w}\|\pi_\text{ref}] \Big\}\Bigg]. \end{aligned}$$7. On Big-Math with group size $$\begin{aligned} J_\text{GRPO}(w) = \mathbb{E}\Bigg[\frac{1}{G} \sum_{g=1}^G \frac{1}{O_g} \sum_{t=1}^{O_g} \Big\{ &\min \Big( \frac{\pi_w(o_{g,t}|a,o_{g,<t})}{\pi_{w_\text{old}}(o_{g,t}|a,o_{g,<t})}, \ &\text{clip}\Big(\frac{\pi_w(o_{g,t}|a,o_{g,<t})}{\pi_{w_\text{old}}(o_{g,t}|a,o_{g,<t})},1-c,1+c\Big)\Big)\,\widehat{A}_{g} \ &\qquad - \mu \,\mathbb{D}_\text{KL}[\pi_{w}\|\pi_\text{ref}] \Big\}\Bigg]. \end{aligned}$$8, about 44% of prompts are silent under the logged distribution because all samples in the group are either correct or incorrect, and therefore contribute zero GRPO signal (Bay et al., 30 Jun 2026). This matters for Iso-GRPO because any “iso” normalization that aims to equalize prompt contributions must contend with the fact that standard GRPO does not merely reduce variance; it reweights learning by disagreement.

A second reinterpretation shows that GRPO is implicitly a process reward model. Under a token-level DAPO objective with one update per batch, the group-relative outcome advantage

$$\begin{aligned} J_\text{GRPO}(w) = \mathbb{E}\Bigg[\frac{1}{G} \sum_{g=1}^G \frac{1}{O_g} \sum_{t=1}^{O_g} \Big\{ &\min \Big( \frac{\pi_w(o_{g,t}|a,o_{g,<t})}{\pi_{w_\text{old}}(o_{g,t}|a,o_{g,<t})}, \ &\text{clip}\Big(\frac{\pi_w(o_{g,t}|a,o_{g,<t})}{\pi_{w_\text{old}}(o_{g,t}|a,o_{g,<t})},1-c,1+c\Big)\Big)\,\widehat{A}_{g} \ &\qquad - \mu \,\mathbb{D}_\text{KL}[\pi_{w}\|\pi_\text{ref}] \Big\}\Bigg]. \end{aligned}$$9

is exactly equivalent to a Monte Carlo step-level reward assignment over shared prefixes. For each process set $$\widehat{A}_{g} = \frac{r_g - \text{mean}(\{r_1,\ldots,r_G\})}{\text{std}(\{r_1,\ldots,r_G\})},$$0 of completions sharing a common prefix, the induced step reward is

$$\widehat{A}_{g} = \frac{r_g - \text{mean}(\{r_1,\ldots,r_G\})}{\text{std}(\{r_1,\ldots,r_G\})},$$1

and the paper proves $$\widehat{A}_{g} = \frac{r_g - \text{mean}(\{r_1,\ldots,r_G\})}{\text{std}(\{r_1,\ldots,r_G\})},$$2 (Sullivan, 25 Sep 2025). In this view, vanilla GRPO weights each process step by $$\widehat{A}_{g} = \frac{r_g - \text{mean}(\{r_1,\ldots,r_G\})}{\text{std}(\{r_1,\ldots,r_G\})},$$3, the number of completions passing through that shared prefix, which can hinder both exploration and exploitation. The proposed $$\widehat{A}_{g} = \frac{r_g - \text{mean}(\{r_1,\ldots,r_G\})}{\text{std}(\{r_1,\ldots,r_G\})},$$4-GRPO divides token losses by $$\widehat{A}_{g} = \frac{r_g - \text{mean}(\{r_1,\ldots,r_G\})}{\text{std}(\{r_1,\ldots,r_G\})},$$5, thereby canceling that multiplicity factor and equalizing step contributions. Empirically, non-trivial overlap is reported to be abundant: with group size 6, only 12 out of 6,700 process trees are trivial, and with group size 36, none of 1,100 trees are trivial (Sullivan, 25 Sep 2025).

Together, these analyses show that “iso” can refer to at least three different normalizations: equalizing across timesteps, equalizing across reward scales, or equalizing across prompt or process-step disagreement. The literature does not collapse these into a single formalism, but it increasingly treats them as closely related design axes.

6. Neighboring variants, empirical profile, and limitations

Several adjacent methods clarify what Iso-GRPO is not, and where its design space remains open. BPPO keeps GRPO’s full-group advantage normalization unchanged,

$$\widehat{A}_{g} = \frac{r_g - \text{mean}(\{r_1,\ldots,r_G\})}{\text{std}(\{r_1,\ldots,r_G\})},$$6

but applies the update only to the shortest correct and shortest incorrect completion in each prompt group, and only to the first $$\widehat{A}_{g} = \frac{r_g - \text{mean}(\{r_1,\ldots,r_G\})}{\text{std}(\{r_1,\ldots,r_G\})},$$7 tokens of those completions (Zhao et al., 27 May 2026). Its contribution is not an “iso” normalization in the temporal or reward-scale sense; rather, it is a sparsification of where group-relative advantages are applied. The paper reports up to $$\widehat{A}_{g} = \frac{r_g - \text{mean}(\{r_1,\ldots,r_G\})}{\text{std}(\{r_1,\ldots,r_G\})},$$8 speedup over GRPO while maintaining competitive accuracy, and mean response-length reductions of approximately 30–50% without an explicit length penalty (Zhao et al., 27 May 2026). This shows that group-relative normalization can coexist with aggressive structural pruning of completions and tokens.

For diffusion LLMs, the main concern is different. “Stabilizing Reinforcement Learning for Diffusion LLMs” argues that directly porting GRPO to dLLMs is unstable because importance ratios are only estimated through noisy ELBO-based proxies. Under these conditions, GRPO’s conditional clipping can be anomalously bypassed by model-agnostic estimation noise, while fixed group-size normalization amplifies gradient-magnitude fluctuations. StableDRL replaces this with unconditional clipping and self-normalization, constraining the update to the convex hull of per-sample gradients (Zhong et al., 6 Mar 2026). This is not called Iso-GRPO, but it addresses the same underlying problem of controlling comparability and scale across grouped samples in a diffusion setting.

Across domains, the empirical record is therefore mixed but coherent. Iso-temporal GRPO in video diffusion is presented as a stable and efficient one-step alternative to full-trajectory GRPO (He et al., 15 May 2026). Reward-invariant GRPO in NCO is presented as a robust baseline-free alternative to rollout baselines and critics (Sepúlveda et al., 9 Jun 2026). Critic-free GRPO in wireless optimization shows that group-relative exploration can cut model size and FLOPs while improving performance (Peng et al., 20 Jan 2026). Process-level reinterpretations and disagreement identities then reveal that these behaviors are governed by hidden normalization choices rather than by a single monolithic principle (Bay et al., 30 Jun 2026, Sullivan, 25 Sep 2025).

The limitations are equally consistent. In video diffusion, iso-temporal grouping alone still leaves training unstable unless TGR removes the time-dependent scaling factor, and noisy or poorly calibrated reward models can still misassign credit (He et al., 15 May 2026). In binary-reward theory, the clean group-standard-deviation identity does not directly extend to non-binary rewards (Bay et al., 30 Jun 2026). In diffusion LLMs, clipping and self-normalization introduce bias even as they suppress collapse (Zhong et al., 6 Mar 2026). Theoretical work on GRPO’s alignment objective also makes clear that the KL term behaves effectively as reverse KL at stationarity, so preserving “iso” invariance in the reward model does not by itself determine the overall alignment behavior (Vojnovic et al., 25 Feb 2025).

In current usage, then, Iso-GRPO denotes a moving frontier rather than a finalized algorithmic standard. Its stable core is GRPO’s group-relative, critic-free optimization. Its distinctive “iso” content depends on which axis is being equalized: timestep, reward scale, disagreement magnitude, or process-step contribution.