Verifiable Geometry Rewards: Rigorous Geometric Reasoning

• Verifiable geometry rewards are a class of supervision signals that provide dense, stepwise feedback by evaluating intermediate reasoning steps rather than only final outcomes.
• They leverage rubric-based scoring, formal subgoal verification, multimodal alignment, and latent-space clustering to enforce logical rigor and reproducibility in geometric proofs.
• By addressing issues such as reward hacking and miracle steps, these rewards enhance trust and accuracy in large language models and multimodal applications for geometric problem solving.

Verifiable geometry rewards are a class of supervision signals for training LLMs and multimodal models to perform geometric reasoning. Unlike standard outcome-based supervision—which rewards only the final answer—verifiable geometry rewards provide dense, granular feedback based on intermediate steps, sub-goals, or structured properties of solutions. These rewards are derived from explicit rubrics, formal subgoal verification, symbolic checks, multimodal alignment, or latent-space clustering, each offering a formally-checked and reproducible standard for correctness and logical rigor in geometry. Their development addresses issues such as reward hacking, miracle steps, and hallucinated deductions, enabling robust process-level verification and improving both trust and accuracy in geometric proof generation and problem solving.

1. Motivation and Taxonomy of Verifiable Geometry Rewards

Traditional outcome-based reinforcement learning rewards—such as assigning a reward of 1 for the correct numeric answer and 0 otherwise—are highly susceptible to reward hacking in mathematical and geometric reasoning. Models may arrive at the correct answer through unsound reasoning, memorization, or single-step “miracle” jumps, leading to severe overestimation of actual reasoning ability. Empirical investigations reveal phenomena such as miracle steps (abrupt, unjustified correct outputs) and solution falsification when only the final answer is considered (Yuan et al., 9 Oct 2025).

Verifiable geometry rewards were developed to overcome these limitations. They are characterized by the explicit, stepwise checking of logical correctness. This is realized through several architectures:

• Rubric-based scoring, where each step is matched against a hand-crafted list of proof desiderata.
• Formal sub-goal verification, converting proofs into chains of subgoals, each numerically or propositionally checkable.
• Structured multimodal comparison, aligning solutions to reference diagrams, textual explanations, or logical forms.
• Intrinsic clustering in latent embedding space, using geometric properties of model activations to derive reward signals without external verifiers.

These methodologies now underpin state-of-the-art practices in geometry reasoning RL, both in pure text and multimodal (diagram-text) domains.

2. Rubric and Subgoal-Based Reward Models

The rubric reward model (RRM) (Yuan et al., 9 Oct 2025) introduces expert-crafted rubrics as fine-grained evaluators for geometric proofs. For each problem, rubric items correspond to desirable properties, such as:

• Axiom/definition usage (e.g., correct application of SAS/ASA/SSS)
• Inference validity (deductions must logically follow from prior steps)
• Diagram consistency (entity labels in text and figure must agree)
• Completeness and justification (no unsupported jumps or missing cases)

Given a chain-of-thought $\tau = (s_1, ..., s_T)$, a rubric scorer evaluates each step $s_t$ on all criteria $c_k$. Binary or graded sub-scores are aggregated:

$$R(\tau) = \frac{1}{\sum_{k=1}^K w_k} \sum_{k=1}^K w_k \left[ \frac{1}{T} \sum_{t=1}^T \mathbf{1}\{s_t \models c_k\} \right] - \lambda\frac{\mathrm{flaw}(\tau)}{T}$$

where $w_k$ is the weight for each rubric criterion, and $\lambda$ penalizes miracle steps and logical leaps. This process-oriented reward provides $0$ to $1$ feedback, enabling policy optimization not just for accuracy but for stepwise rigor.

The sub-goal verifiable reward (SGVR) framework (Chen et al., 8 Jan 2026) formalizes process supervision via a sequence of numeric subgoals $(\mathcal{T}_t, y_t)$ derived from formally verified skeletons of geometry proofs. Each subgoal is independently checked, and the Skeleton Rate (SR) is computed as:

$$\mathrm{SR} = \frac{1}{n_i} \sum_{t=1}^{n_i} \mathbb{I}(\hat{y}_{i,t} = y_{i,t})$$

rewarding partial correctness in long reasoning chains and enabling dense RL feedback. This method significantly narrows the gap between final-answer accuracy and genuine reasoning quality.

3. Multimodal and Structured Reward Architectures

The alignment of diagram and text is critical in geometry. StructVRM (Zhang et al., 7 Aug 2025) and GeoVLMath (Guo et al., 13 Oct 2025) extend verifiable reward to multimodal geometry tasks.

StructVRM employs a learned verifier that segments model output into $k$ sub-answers and scores each against reference solutions via:

• Symbolic mathematical equivalence (with computer algebra)
• Semantic equivalence (thresholded cosine similarity of sentence embeddings)

The structured reward is:

$$R_{\text{struct}}(\hat{y}, y) = \frac{1}{k} \sum_{j=1}^k s_j$$

where $s_j$ reflects mathematical or semantic equivalence on sub-question $j$. This architecture allows partial credit and efficiently captures the correctness of complex, multi-step solutions.

GeoVLMath uses a cross-modal reward model $R_{\phi}$ trained to score the agreement between a generated auxiliary-line description and an annotated ground-truth diagram:

$r_{\text{aux}} = R_{\phi}(I, d, I^+) \in [0, 1]$

where $I$ is the original diagram, $d$ is the textual auxiliary-line construction, and $I^+$ is the diagram with the auxiliary line(s). The composite RL reward is a convex combination of auxiliary-line alignment and final-answer accuracy, supporting precise diagram-text alignment and rigorous geometric justification (Guo et al., 13 Oct 2025).

4. Intrinsic Latent Geometry and Self-Verification

A recent advance is the use of intrinsic geometric structure in the latent space of LLMs for verifiable rewards (Zhang et al., 13 Jan 2026). The Iterative Robust Centroid Estimation (IRCE) procedure exploits the finding that the terminal hidden states of correct reasoning trajectories form dense, well-separated clusters in $\mathbb{R}^d$, while incorrect trajectories are scattered.

The reward for each trajectory is computed by projecting latent vectors onto the unit hypersphere, estimating a robust centroid $\mu$, and scoring:

$$r_i = -\|\tilde{h}_i - \mu\|_2 \quad\Longrightarrow\quad R_i = \frac{r_i - \min_j r_j}{\max_j r_j - \min_j r_j}$$

This self-verifying, continuous reward delivers dense, computationally efficient feedback entirely without external labeling or symbolic checkers, while maintaining empirical performance and robustness in geometry and broader reasoning tasks.

5. Practical Frameworks: Reasoning Gym and Contrastive Policy Objectives

Reasoning Gym (Stojanovski et al., 30 May 2025) operationalizes verifiable rewards via a suite of procedural geometry environments. For each instance, a verifier function parses the model’s answer and performs deterministic, programmatic checks using ground-truth metadata (e.g., coordinate comparisons, triangle centers, angle measures). The standard reward is:

$R = r_{\text{acc}} + \alpha r_{\text{fmt}}$

where $r_{\text{acc}}$ is 1 for a correct answer and $r_{\text{fmt}}$ is a small bonus for correct formatting. All verification is automated and reproducible, ensuring full transparency and fidelity.

GeometryZero (Wang et al., 8 Jun 2025) extends this principle using Group Contrastive Policy Optimization (GCPO), which introduces contrastive masking to adaptively reward or penalize auxiliary construction, depending on empirical utility, and incorporates a length-based reward to encourage detailed reasoning chains. All rewards are verifiable by objective code compilation or string matching.

Framework	Reward Source	Verification Mechanism
RRM (Yuan et al., 9 Oct 2025)	Rubric/step criteria	LLM/auditor scoring + flaw penalty
SGVR (Chen et al., 8 Jan 2026)	Numeric subgoals	Symbolic evaluation/indicator
StructVRM (Zhang et al., 7 Aug 2025)	Multimodal sub-question	Symbolic and semantic checks
GeoVLMath (Guo et al., 13 Oct 2025)	Cross-modal auxiliary lines	Model-based score on diagram-text
Latent-GRPO (Zhang et al., 13 Jan 2026)	Embedding geometry	Centroid distance, unsupervised
Reasoning Gym (Stojanovski et al., 30 May 2025)	Numeric/formatting	Verifier algorithm (code)
GeometryZero (Wang et al., 8 Jun 2025)	Group contrast/auxiliary	Compilation, string check

6. Empirical Results and Impact on Geometric Reasoning

Empirical validation across multiple benchmarks demonstrates pronounced improvements:

• RRM in geometry boosts Verified Pass@512 from 19.0% (outcome-only) to 41.8%, and slashes miracle step rates from 58.7% to 16.9% on representative 4B models. Diagram-consistency rubrics are especially vital; ablation cuts verified accuracy from 41.8% to 31.4% (Yuan et al., 9 Oct 2025).
• SGVR increases average geometric reasoning accuracy by +9.7 points and improves subgoal skeleton rates from 50.2% to 87.7% (Chen et al., 8 Jan 2026).
• StructVRM delivers gains on the “Math” track of STEM-Bench (83.26% → 86.15%) and enables rigorous, partial-credit supervision irrespective of proof format (Zhang et al., 7 Aug 2025).
• GeoVLMath outperforms much larger vision-LLMs on auxiliary-line benchmarks, e.g., +10.19% (26.12% vs 15.93% Pass@5) on the GeoAuxBench “Hard” tier (Guo et al., 13 Oct 2025).
• Latent-GRPO/IRCE achieves ∼2× speed-up with equal or improved accuracy versus external verifiers, with GSM8K accuracy of 73.88% (vs 64.20% LLM-Judge) (Zhang et al., 13 Jan 2026).
• GCPO in GeometryZero yields up to +4.23% over previous GRPO methods, with explicit verifiability at every step (Wang et al., 8 Jun 2025).

7. Best Practices, Pitfalls, and Verification Protocols

Best practices include:

• Craft rubrics and subgoal mappings to be method-agnostic and avoid overfitting to specific solution styles (Yuan et al., 9 Oct 2025).
• Incorporate explicit failure-mode checks for miracle steps, unsupported leaps, and off-topic steps.
• Calibrate LLM-based or model-based verifiers on human-annotated proofs, auditing for correlation between reward and logical soundness.
• For contrastive/auxiliary construction rewards, use empirical masking to reward only those behaviors that demonstrably enhance accuracy (Wang et al., 8 Jun 2025).
• Automated verification, code compilation, and deterministic checkers are preferred for large-scale RL to ensure reproducibility and eliminate dependence on subjective human evaluation (Stojanovski et al., 30 May 2025, Wang et al., 8 Jun 2025).
• For intrinsic geometry-based rewards (e.g., IRCE), routinely monitor clustering diagnostics (silhouette scores, Dunn index, rank-correlation with external judges) to maintain reward fidelity under distributional shifts (Zhang et al., 13 Jan 2026).

A plausible implication is that verifiable geometry rewards constitute an extensible paradigm not just for geometry, but for general mathematical, logical, and multimodal reasoning; their architecture is readily adapted to new domains with formal annotators and structured verifiers.

---

In summary, verifiable geometry rewards comprise a rigorous, process-supervised approach to model training for geometric reasoning, replacing sparse outcome signals with structured, stepwise, and fully auditable signals. This paradigm underpins the leading models and benchmarks in geometric RL, continually advancing both the trustworthiness and the accuracy of automated mathematical reasoning (Yuan et al., 9 Oct 2025, Zhang et al., 7 Aug 2025, Guo et al., 13 Oct 2025, Chen et al., 8 Jan 2026, Zhang et al., 13 Jan 2026, Wang et al., 8 Jun 2025, Stojanovski et al., 30 May 2025).