ECCV2026 Segmentation Multimodal VLM reasoning segmentation reinforcement learning GRPO test-time scaling parameter-efficient fine-tuning

StAR: Segment Anything Reasoner¶

Conference: ECCV2026
Paper: ECCV
Code: https://github.com/ysj9909/StAR
Area: Segmentation / Multimodal VLM
Keywords: reasoning segmentation, reinforcement learning, GRPO, test-time scaling, parameter-efficient fine-tuning

TL;DR¶

Rather than proposing a new architecture, StAR dismantles the bottlenecks that suppress a base model's reasoning ability in an existing RLVR framework (VisionReasoner) along four axes—parameter-tuning scheme, reward design, learning strategy, and answer format—and is the first to bring parallel test-time scaling (mask-level majority voting) to segmentation; with only 5k training samples it lifts Qwen2.5-VL 7B on the ReasonSeg-X test set from 42.2 to 49.2 gIoU (50.3 with majority voting), while also contributing a cleaner ReasonSeg-R and a harder ReasonSeg-X benchmark.

Background & Motivation¶

LISA was the first to formalize "implicit query + image → target mask" as reasoning segmentation, and most work since follows the same recipe: wire the reasoning ability of an MLLM to the mask generation ability of the SAM family, where the MLLM reads the query, performs visual-linguistic reasoning, and emits boxes/representative points that SAM turns into masks. VisionReasoner, SAM-R1, Seg-Zero and others went further by introducing Reinforcement Learning with Verifiable Rewards (RLVR), using GRPO to strengthen the MLLM's ability to localize through reasoning. Among them, VisionReasoner offers a clean decoupled "reasoning module–mask generator" framework plus a full set of Hungarian-matching multi-object rewards, and has become the de facto baseline of this line.

Yet one question has been broadly skipped: does the current RLVR framework actually surface the visual reasoning potential that the base model does not readily manifest? The authors' observation is that it does not, and that the loss is spread across every pillar. On the parameter side, full-parameter tuning rewrites all weights, doubling memory (an extra reference model is needed for the KL penalty) and easily damaging the base model's world knowledge. On the reward side, accuracy rewards only track geometric precision of boxes and points, misaligned with the task's real endpoint—mask quality—and the prevailing hard reward ("1 if IoU exceeds a threshold, otherwise 0") neither reflects incremental improvement during training nor avoids reward hacking. On the learning side, vanilla GRPO collapses to near-zero reward variance within a group on samples that are too easy or too hard, so the advantage signal vanishes—exactly on the hard problems where sparse positive feedback is most likely to elicit the base model's "latent but hard to elicit" reasoning skills. On the output side, requiring only geometric coordinates quietly turns this semantics-first task into a coordinate-guessing mode, while long CoT progressively decays attention over visual tokens and amplifies hallucination. Evaluation is no more solid: parts of ReasonSeg have poor mask quality or contain reasoning errors, its reasoning depth is limited, and it never defines which reasoning types it covers, making systematic evaluation of frontier-MLLM methods difficult.

This paper's angle is not a new architecture but a new recipe: retrofit each pillar of RLVR in a computationally efficient way, and bring in the previously ignored parallel test-time scaling—majority voting is natural for tasks with a discrete answer space but does not hold for pixel-level, multi-object prediction, so it needs dedicated design. Core idea: keep rather than rewrite the base model's knowledge via high-rank LoRA with a larger learning rate and no KL penalty; pull SAM into the RL loop and use a tiered mask-IoU reward to provide fine-grained, task-aligned signals, then use REST to expand the rollout pool to 256 and update only on the two advantage extremes; re-anchor localization in semantics through a label-before-coordinates answer format; and finally apply mask-level majority voting (IoU clustering plus vote counts) at test time for parallel test-time scaling.

Method¶

Overall Architecture¶

StAR keeps the decoupled reasoning–segmentation forward path: given an image and an implicit query, the MLLM first produces a chain-of-thought trace, then emits an explicit semantic label along with the target's bounding box and representative point; these geometric outputs serve as visual prompts to a frozen SAM 2 Large, which returns the final binary masks. Only the MLLM is updated; SAM 2 never is. The paper is therefore really about two recipes for one fixed forward path: a training recipe that retrofits GRPO along four design axes, and a test recipe that aggregates multiple parallel responses at the mask level.

It is worth stating up front how the reasoning trace and the mask promote each other, because it determines why every later design looks the way it does. StAR takes a single-directional reason-then-segment route: the CoT and the semantic label determine the geometric prompt, the geometric prompt determines the mask, and the model never looks back at the mask to revise its reasoning. The genuine two-way coupling happens inside the training loop: the mask SAM produces is scored into a mask-IoU reward, which joins the MLLM-level accuracy rewards to form the GRPO advantage, and the gradient flows back into the MLLM—so "how well it segmented" is translated into a learning signal about "how well it reasoned." In other words, the reasoning trace proposes the target and its geometric prior, mask quality scores that trace, and the two shape each other through the RLVR reward loop; segmentation is not a separately trained module here but the verifiable outcome of reasoning.

The relationship to the SAM family should be explicit: SAM 2 is neither fine-tuned nor invoked as an external tool across multiple agentic turns—it is a fixed, prompt-in/mask-out component of the forward path, and its mask quality score is additionally reused as the confidence for test-time aggregation. What is evaluated is reasoning segmentation itself (implicit query → mask, measured by gIoU/cIoU), not some reasoning-assisted segmentation side task. Compared with SAM 3's agentic pipeline centered on sequential refinement, StAR's test-time scaling sits on the opposite axis: sample in parallel, then aggregate, rather than refine round by round.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["implicit query + image"] --> B["high-rank LoRA policy<br/>frozen base, low-rank deltas only"]
    B --> C["label prediction<br/>name the target, then emit box/point"]
    C --> D["SAM 2 (frozen)<br/>visual prompt → candidate masks"]
    D --> E["tiered mask-IoU reward<br/>+ MLLM-level accuracy rewards"]
    E --> F["REST<br/>16 → 256 rollouts, update both extremes"]
    F -->|GRPO iterations| B
    D -->|32 parallel samples at test time| G["mask-level majority voting<br/>IoU clustering + vote counts"]
    G --> H["final masks"]

Key Designs¶

1. High-rank LoRA instead of full-parameter tuning: RLVR needs to learn how to use knowledge, not to rewrite it

The most counterintuitive move in this paper is deliberately replacing full-parameter tuning with LoRA. The pain point is specific: in SFT, full-parameter tuning exists to densely transfer new knowledge, but the principle for applying RLVR to reasoning segmentation is the opposite—preserve the base model's knowledge and reasoning patterns and only learn how to leverage them for segmentation. Existing RLVR segmentation methods (and most open-source reasoning models) nevertheless copy full-parameter tuning wholesale, which forces an extra reference model in memory for the KL penalty and rewrites weights broadly enough to damage the base model's world knowledge. LoRA models sparse updates with a low-rank decomposition and therefore largely preserves that knowledge, but the default configuration drops noticeably, so two adjustments are needed: raise the learning rate from \(1\times10^{-6}\) to \(1\times10^{-5}\) to speed up task adaptation and remove the KL penalty term, then raise the rank from 16 to 64 to widen the expressiveness of the update matrices so they can carry more complex reasoning.

The outcome is that this "cheaper" configuration is also stronger: ReasonSeg-X rises from 42.2% to 45.4% and ReasonSeg-R from 64.8% to 66.1%, while training cost falls to less than half of VisionReasoner's. The saved compute is not wasted—it is reinvested into rollout scaling (design 4), which is exactly the resource-reallocation logic the paper keeps returning to: low rank frees memory, memory buys more samples.

2. Label prediction: name the target before emitting coordinates

Reasoning segmentation is semantic-first: the model must understand the query and complete compositional cross-modal inference before it knows what to enclose, yet the deliverable is geometric. This semantic-in/geometric-out asymmetry quietly biases generation toward a coordinate-only mode in which the model emits plausible-looking boxes without ever committing to what it is enclosing. Worse, prior work shows that as CoT grows longer, attention over visual tokens decays markedly and hallucination increases. StAR's countermeasure is almost too small to call a design: add a label field to the answer schema and require a short semantic naming before the coordinates, e.g. {"label": "the owner of the largest dog", "bbox_2d": [...], "point_2d": [...]}. The change touches only the prompt/response format and no parameters, yet it systematically reshapes the generation dynamics by inserting an explicit semantic "naming" step immediately before coordinate prediction. The predicted labels tend to paraphrase the query or even answer it, semantically anchoring the predicted region back to the query and improving reasoning faithfulness; measured on StAR-7B, the attention mass over visual tokens during coordinate prediction rises from 5.3% to 6.3%.

The ablation shows that what matters is the order, not the label itself: after training with label-first, reversing the order to label-last at test time degrades performance sharply; conversely, for a model trained label-last, prompting label-first at test time improves performance substantially. The benefit comes from reformulating localization as a semantically-conditioned coordinate prediction step in which the already-emitted label text anchors the subsequent grounding.

3. Tiered mask-IoU reward: pull SAM into the RL loop so rewards align with the task endpoint

VisionReasoner's accuracy rewards look only at geometric precision of boxes and points, which is misaligned with the task's endpoint (the mask) and invites reward hacking (the left side of Fig. 5 shows a typical misalignment); the hard "1 above an IoU threshold, 0 otherwise" reward neither reflects incremental improvement nor enables fine-grained evaluation. Following SAM-R1, StAR explicitly brings SAM into the RLVR training loop and uses a tiered mask-IoU reward: mask quality is discretized into several IoU bands, with high bands scoring high and low quality scoring low, yielding a graded signal:

\[ r_{\text{mask}}=\begin{cases} 5, & \text{IoU} > 0.90\\ 1, & 0.30 < \text{IoU} \le 0.90\\ 0.6, & \text{otherwise} \end{cases} \]

(⚠️ The typesetting of this equation is corrupted in the cached PDF; the exact correspondence between tier values and thresholds should be checked against the original. What is legible are the thresholds 0.90 and 0.30 and the tier values 5 / 1 / 0.6.) This reward is used alongside MLLM-level rewards: a bbox IoU reward assigns 1 when the predicted box exceeds 0.5 IoU with the GT box, and a bbox/point L1 reward assigns 1 when the L1 distance is below 10 / 30 pixels respectively, plus VisionReasoner's original Thinking / Answer / Non-repeat format rewards. Multi-object matching still uses a batched Hungarian algorithm to solve the assignment between \(N_{\text{pred}}\) predictions and \(N_{\text{GT}}\) instance annotations, and the resulting accuracy reward is divided by \(\max\{N_{\text{pred}}, N_{\text{GT}}\}\) to suppress over- and under-segmentation.

Keeping the MLLM-level rewards strengthens coarse localization and penalizes reasoning failures more heavily, while the tiered mask reward acts as a curriculum-like signal: early training focuses on locating the target through high-quality reasoning, and later stages shift to cutting the target boundary accurately. The most telling evidence is rollout utilization: on VisionReasoner, raising rollouts from 8 to 16 brings essentially nothing, whereas StAR's finer-grained reward discriminates rollouts well enough that the same move yields a real gain (45.4% → 47.1% / 66.1% → 66.6%), with overall training cost still modest.

4. REST: expand the rollout pool to 256 and update only the advantage extremes

Stage-2 GRPO training has a hidden failure mode: advantage vanishing. For samples that are too easy or too hard, rewards within the group are nearly identical, the normalized variance approaches zero, and the gradient signal disappears; in practice, on average more than half of the batch sits in this near-zero-variance regime during vanilla GRPO training—yet it is precisely the sparse positive feedback on very hard problems that is most likely to surface the base model's dormant reasoning skills. REST (Rollout-Expanded Selective-Tuning), a drop-in modification to GRPO, attacks this by decoupling exploration from learning: first expand the rollout pool from the conventional small \(n\) (8–16) to \(N\)=128–256 to explore a far wider space of reasoning trajectories, then skip the full update and instead perform advantage-extremal selection, updating only on an informative subset of size \(m\)—the \(m/2\) rollouts with the largest advantages plus the \(m/2\) with the smallest. Taking both tails aligns with reward variance-maximization strategies and yields stronger contrastive signals.

REST is cheap because it exploits the compute/memory asymmetry of LLM RL: rollout generation is highly parallel, needs minimal activation storage, and is high-throughput when batched, so extra samples amortize well; policy updates, by contrast, are memory- and communication-intensive because of optimizer states and cross-device synchronization. Exploring a 16× larger space (\(n\) from 16 to 256) therefore raises end-to-end wall-clock time by only about 2× (from 170 s to 364 s per step), leaving the expensive part nearly constant. Effectively, REST raises the chance of observing rare successful reasoning paths on hard prompts and internalizes them; with only the 240 training samples of ReasonSeg-X, Stage-2 training adds another 1.5 / 1.6 points on ReasonSeg-X / ReasonSeg-R.

5. Mask-level majority voting: move majority voting down to the pixel level

Majority voting is well-defined for a discrete answer space (e.g. checking whether two numeric answers are equal) but breaks down on pixel-level, multi-object prediction—two masks cannot "vote for equality." StAR instead pools all mask candidates from the parallel responses and clusters them greedily by IoU: for each candidate, compute its IoU against the representative mask of existing clusters; above a threshold of 0.85, assign it to that cluster, otherwise start a new one. The number of target instances \(\hat{K}\) is the mode of the per-response predicted target counts, and if no-target prediction is the majority the output is simply "no target object." Clusters are then ranked by vote count, the top \(\min(\hat{K}, \#\text{clusters})\) are selected, and within each selected cluster the single instance mask with the highest SAM mask quality score is kept. The overall prediction is the union of the selected instance masks. The strategy is framework-agnostic—any MLLM-based segmentation framework can adopt it—and on StAR's own wide reasoning-pattern distribution (32 parallel samples by default) it improves nearly every benchmark and every model scale, letting the 7B model match the overall score of the 72B SAM 3 Agent on the ReasonSeg-X test set.

A Worked Example¶

Take the receipt sample in Fig. 7, whose query asks for the group of menu-item lines marking the first point at which the diners' orders diverge after everyone initially ordered the same item. At training time the flow runs like this: the model expands a CoT—counting row occurrences, finding that "Frangelico" appears four times and was therefore ordered by everyone, then noticing that "Espresso" and "Cappuccino" appear in differing combinations, and concluding that the divergence starts there. It then follows the label-prediction format, emitting {"label": "divergent menu-item group", ...} first and the corresponding bbox and representative point afterwards. The geometric prompt goes into the frozen SAM 2, the resulting mask is compared with the GT mask, and the IoU lands in a middle tier, far from saturated. REST acts here: among 256 rollouts, a few get the divergence point right and cut the mask precisely (IoU > 0.90, top tier) and form the positive extreme, while a batch that encloses the wrong region with IoU near zero forms the negative extreme; only the \(m\)=16 rollouts at these two extremes drive the LoRA policy update, turning "how to count and how to compare" into gradient.

At inference the same query is sampled 32 times in parallel and each response's mask candidates are clustered greedily at an IoU threshold of 0.85: most responses enclose the same menu-item lines and form one high-vote cluster, while the occasional candidate that swallows the whole receipt has insufficient IoU with the representative mask and forms its own low-vote cluster. The mode of the predicted target counts gives \(\hat{K}\); the top \(\hat{K}\) clusters by votes are taken, the best mask in each cluster is chosen by SAM's mask quality score, and their union is the final output—more stable than trusting the first response alone.

Loss & Training¶

Training proceeds in two stages, both using GRPO. Stage 1 aims to elicit flexible localization while maximally preserving pretrained visual perception, so it uses VisionReasoner's collected training set with the reasoning samples (LISA++) removed, keeping only LVIS, RefCOCOg, and gRefCOCO for 5k samples in total. Stage 2 continues fine-tuning the Stage-1 model on the ReasonSeg-X training set (240 samples) to activate complex reasoning potential and produce the final model. GRPO itself keeps the group-normalized advantage:

\[ A_i=\frac{r_i-\text{mean}(\{r_i\}_{i=1}^{n})}{\text{std}(\{r_i\}_{i=1}^{n})},\qquad \mathcal{J}_{\text{GRPO}}(\theta)=\mathbb{E}\Big[\frac{1}{n}\sum_{i=1}^{n}\min\big(\rho_i A_i,\ \text{clip}(\rho_i,1-\epsilon,1+\epsilon)A_i\big)-\beta\,\mathbb{D}_{\text{KL}}(\pi_\theta\|\pi_{\text{ref}})\Big] \]

where \(\rho_i=\pi_\theta(o_i|q)/\pi_{\theta_{\text{old}}}(o_i|q)\) and \(\epsilon,\beta\) are hyperparameters; StAR's LoRA configuration removes the KL penalty (i.e. \(\beta=0\)), which is also why it can drop the extra reference model. In practice the base models are Qwen2.5-VL 7B and Qwen3-VL 8/32B, with SAM 2 Large as the mask generator; batch size is 16, the learning rate is \(1\times10^{-5}\) (\(5\times10^{-6}\) for Stage 2), LoRA rank 64 is used for all base models, REST is enabled only in Stage 2, and majority voting samples 32 reasoning paths by default.

Key Experimental Results¶

Main Results¶

The metrics are gIoU and cIoU: gIoU averages per-sample IoU, while cIoU divides cumulative intersection by cumulative union; the former weights every sample equally and the latter is more sensitive to large targets, so reporting both avoids cherry-picking the more favorable one. Benchmarks are ReasonSeg (RS), the authors' cleaned ReasonSeg-R, the new ReasonSeg-X (four reasoning types: P/F, C/KI, C/R, C/MH), and MMR, plus RefSpatial-Bench for robotic spatial referring.

Method	Base model	RS test gIoU/cIoU	RS-R gIoU/cIoU	RS-X test gIoU/cIoU	RS-X C/MH gIoU/cIoU
LISA (ft)	Llama2 13B	51.5 / 51.3	52.5 / 53.3	25.1 / 26.0	13.8 / 16.8
SAM-R1	Qwen2.5-VL 7B	60.2 / 54.3	–	–	–
SAM-Veteran	Qwen2.5-VL 7B	62.6 / 56.1	–	–	–
VisionReasoner	Qwen2.5-VL 7B	63.6 / 55.7	64.8 / 56.8	42.2 / 33.8	24.5 / 23.8
StAR stage-1	Qwen2.5-VL 7B	66.7 / 60.8	69.0 / 65.6	47.4 / 40.6	24.5 / 25.1
SAM 3 Agent (sequential refinement)	Qwen2.5-VL 72B	71.8 / 65.2	72.4 / 65.3	49.8 / 40.3	30.8 / 35.1
StAR	Qwen2.5-VL 7B	67.5 / 61.3	69.7 / 66.2	49.2 / 43.6	28.0 / 28.8
StAR + MV	Qwen2.5-VL 7B	68.5 / 65.0	70.7 / 67.2	50.3 / 44.9	27.9 / 30.5
StAR + MV	Qwen3-VL 8B	71.8 / 66.5	74.9 / 69.7	59.6 / 53.1	39.5 / 41.7
StAR + MV	Qwen3-VL 32B	72.7 / 68.1	75.0 / 70.6	64.1 / 60.8	47.9 / 50.3

Generalization across other benchmarks:

Benchmark	Setting	StAR	Compared against	Gap
MMR Obj&Part gIoU	zero-shot	33.0 (7B + MV)	28.4 (VisionReasoner 7B)	+4.6
MMR Obj test gIoU	zero-shot	45.6 (7B + MV)	37.2 (VisionReasoner 7B)	+8.4
MMR Part gIoU	zero-shot	14.9 (7B + MV)	13.6 (M2SA, Llama2 13B, trained on MMR)	+1.3
RefSpatial-Bench Location	zero-shot	62.00 (8B)	52.00 (RoboRefer-8B)	+10.00
RefSpatial-Bench Unseen	zero-shot	38.96 (8B)	37.66 (RoboRefer-8B)	+1.30

Ablation Study¶

Design trajectory (retrofitting along the four axes, +7.0 / +4.9 gIoU on ReasonSeg-X/R in total):

Config	RS-X test gIoU	RS-R gIoU	Note
VisionReasoner-7B (start)	42.2	64.8	full-parameter tuning + original rewards + geometric-only answers
+ high-rank LoRA (r16 → 64, LR 1e-5, no KL)	45.4	66.1	training cost below half of the baseline
+ tiered mask-IoU reward, rollout 8 → 16	47.1	66.6	rollout scaling only pays off once the reward is finer
+ Stage-2 (REST, n 16 → 256, m 16)	48.6	68.2	another 1.5 / 1.6 from just 240 samples
+ label prediction (final StAR-7B)	49.2	69.7	visual attention mass 5.3% → 6.3%

REST exploration strength and answer order (RS-X test gIoU; the default setting is shown in bold):

REST samples n	StAR-7B	StAR-7B + MV	StAR-8B	StAR-8B + MV	Per-step time
16	48.4	49.3	57.1	58.1	170 s
64	49.2	49.5	56.9	58.3	214 s
128	49.4	49.8	57.3	59.0	260 s
256	49.2	50.3	57.9	59.6	364 s

Answer order	StAR-7B RS-R	StAR-7B RS-X	StAR-8B RS-R	StAR-8B RS-X
label first (default)	69.7	49.2	73.8	57.9
label last	68.2	46.3	73.1	55.8

Key Findings¶

Stage 1 already overtakes every same-backbone competitor: without touching any reasoning data and doing RLVR on just 5k referring-segmentation samples, StAR stage-1 reaches 47.4 gIoU on the ReasonSeg-X test set (VisionReasoner: 42.2), showing that the first four design axes pay off mainly through "not damaging the base model plus aligning the reward with the mask" rather than through piling on reasoning data.
Finer rewards are the precondition for rollout scaling: raising rollouts from 8 to 16 yields nothing on VisionReasoner, while the same move on StAR, with its tiered mask reward, brings +1.7 / +0.5. More sampling is not a free lunch—it converts into gradient only when the reward can discriminate between rollouts.
The model-size × benchmark-difficulty interaction is informative: StAR-8B and StAR-32B are nearly tied on ReasonSeg-R (74.9 vs 75.0), but the gap opens immediately on ReasonSeg-X with its heavier C/R and C/MH share (59.6 vs 64.1). Larger models hold more suppressed reasoning potential, and a deep-reasoning benchmark like ReasonSeg-X is necessary to expose it.
What matters in label prediction is order, not the label: the same trained model, moved to label-last at test time, drops from 49.2 to 46.3 on RS-X; conversely, a model trained label-last improves noticeably when prompted label-first. The gain comes from the reformulation into semantically-conditioned coordinate prediction, not from the extra text.
Majority voting gains scale with exploration and base capability: MV adds about +1.1 RS-X gIoU at 7B, +1.7 at 8B, and +2.4 at 32B; raising REST's \(n\) from 16 to 256 moves the MV gain from +0.9 to +1.1 at 7B. The authors attribute this to the training and test sampling distributions being brought closer together.

Highlights & Insights¶

Saving compute and gaining accuracy are the same move here: the memory freed by LoRA is explicitly reinvested into rollout expansion, and rollouts are the cheap, parallel part—REST buys a 16× larger exploration space for only about 2× wall clock. Splitting training compute into "expensive" and "cheap" buckets and reallocating between them transfers directly to any GRPO-style training.
One extra field in the answer format can change attention allocation: label prediction only alters the prompt/response schema, yet it pushes visual-token attention mass from 5.3% to 6.3%. Compared with methods that restructure training or inference to strengthen visual dependency, this is an extremely low-cost alignment lever worth trying in other grounding-heavy tasks.
Treat the reward function as the interface of verifiability: SAM is frozen and never trained, yet its output becomes a reward signal, letting a non-differentiable segmentation component supervise a differentiable reasoning model. This "frozen tool → measure → reward" pattern generalizes to any task whose reasoning output is consumed by an expensive tool.
The longest axis turned out to be parallel, not sequential: on the axis opposite to SAM 3's agentic refinement, the paper shows the ceiling for parallel sampling is far from reached, and that stronger and larger base models benefit more.

Limitations & Future Work¶

The authors' own limitations sit on the evaluation and data side: ReasonSeg-X has only 1,169 samples, of which just 240 are for training, so Stage-2 gains hinge heavily on the quality and coverage of those 240; the boundary between P/F and C/KI reasoning types is admittedly blurred, and while the authors say they separated them strictly, no annotator-agreement figure is reported.
The cost of test-time scaling is presented generously: MV samples 32 paths by default, and while the paper reports gains, a 32× inference overhead is not always acceptable in real deployment (robotics in particular), and no full accuracy/latency curve against the number of MV samples is given.
The mask reward is a hand-tuned piecewise function; why the thresholds are 0.90 and 0.30, and why the tiers take the values 5 / 1 / 0.6, receive no sensitivity analysis (⚠️ and the equation's typesetting is corrupted in the cached PDF, so the tier details should be checked against the original). A poorly chosen tiering could reintroduce reward hacking.
Improvement directions: making REST's \(m\) and \(n\) adaptive over the course of training (wider early, narrower later), and making MV's 0.85 clustering threshold adaptive to prediction uncertainty rather than fixed, are both low-risk, high-return next steps.

vs VisionReasoner: On the same decoupled "MLLM emits boxes/points + SAM emits masks" route, VisionReasoner uses full-parameter tuning, geometry-only rewards, and a geometric-only answer format; StAR swaps these for high-rank LoRA, a tiered mask reward plus label prediction, and adds Stage-2 training on its own 240 samples. At the same backbone and scale, ReasonSeg-X test goes from 42.2 to 49.2—the most direct comparison in the paper.
vs SAM-R1: SAM-R1 was first to bring SAM into the RLVR reward, and StAR explicitly marks its tiered mask reward as an addition on top of it (extra tiers). The difference is that StAR also overhauls the parameter scheme, the RL strategy, and the answer format, and adds test-time scaling—its contribution is closer to a systematic sweep of the whole design space than a single-point reward improvement.
vs SAM 3 Agent: SAM 3 Agent takes the sequential-refinement route and reaches 49.8 RS-X gIoU with a 72B backbone; StAR-7B with parallel majority voting reaches 50.3, showing that for a task with a continuous output space, parallel test-time scaling is a route worth taking seriously alongside sequential refinement—and the two axes should in principle stack.
vs LISA: LISA established the task formulation and the ReasonSeg benchmark, but the benchmark defects it exposed (mask quality, reasoning errors, undefined reasoning types) are exactly what ReasonSeg-R/X answer. Moving from "proposing the task" to "making the evaluation solid" is also a sign of a maturing field.

Rating¶

Novelty: ⭐⭐⭐⭐ Individual techniques are mostly recombinations of existing ideas (LoRA, tiered rewards, majority voting), but systematizing the RLVR design space and being the first to bring parallel test-time scaling to segmentation makes the overall framework new.
Experimental Thoroughness: ⭐⭐⭐⭐⭐ Four reasoning-segmentation benchmarks plus robotic spatial referring, three base-model scales (7B/8B/32B), and ablations covering every design axis with training-time cost.
Writing Quality: ⭐⭐⭐⭐ The design trajectory (Fig. 4) is laid out clearly and motivations map one-to-one onto ablations; the weak point is that some equations and figures are corrupted in the PDF, so tier details require checking the original.
Value: ⭐⭐⭐⭐⭐ For teams working on RLVR plus dense prediction, this is a directly reusable checklist of which configurations drag performance down, and ReasonSeg-R/X will be used by follow-up work for a long time.