Reinforcement Learning with Verifiable Rewards
RLVR
Practice with an answer checker.
The model tries a task. A checker scores the result. Training makes higher-scoring attempts more likely.
The idea
You practice arithmetic with an answer key. Nobody needs to vote on whether 6 × 7 equals 42. But a tiny answer key cannot test every skill you might need.
The feedback comes from a check, not a vote.
Step 1 of 6
Start with a question that has a checkable answer.
Each training row stores a question and its correct final answer. The model sees only the question. The answer key stays with the checker, so the model cannot copy it.
Step 2 of 6
The model tries four times.
The same model answers the same question four times. The attempts differ because the model picks each next word with a little randomness. Each finished attempt is called a rollout.
Step 3 of 6
A program checks each final answer.
Nobody votes. The checker reads the text after “Answer:” and compares it with the answer key. A match earns 1 point. Anything else earns 0.
Step 4 of 6
Every attempt now has a reward.
A and C earned 1. B and D earned 0. Look at C: it only guessed, yet it still scored 1, because the checker looks only at the final answer. A reward proves only what the checker tested.
Step 5 of 6
Training makes rewarded answers more likely.
An RL algorithm such as GRPO or PPO nudges the model’s settings. Answers that earned 1 become more likely next time, and answers that earned 0 become less likely. The dashed line marks the chance before the update.
Step 6 of 6
Practice on the next question, then test on new ones.
The updated model, now version 2, gets the next question and the loop repeats thousands of times. At the end it is tested on questions it never practiced, to check that it learned the math and not just these answers.
Question math-0042math-0043
A baker puts 6 rows of 7 cookies on a tray. How many cookies are on the tray?A train travels 60 km per hour for 3 hours. How far does it go?
Answer key
42180 hidden
Model version 1version 2
Reads the question and writes a full answer.
Attempt A
6 × 7 = 42. Answer: 42
reward 1Attempt B
6 + 7 = 13. Answer: 13
reward 0Attempt C
Just a guess. Answer: 42 lucky guess
reward 1Attempt D
6 × 8 = 48. Answer: 48
reward 0 Checker program final answer == 🔒 42 ?
“42” == “42”
“13” ≠ “42”
“42” == “42”
“48” ≠ “42”
Update · chance of each final answer
before the updateafter, rewardedafter, not rewarded
Task or dataModel being trainedJudge or scorerGood / chosenBad / rejectedTraining update
jsontrain.jsonl · one rownew at this step
{
"id": "math-0042",
"prompt": "A baker puts 6 rows of 7 cookies on a tray. How many cookies are on the tray? End with 'Answer: <number>'.",
"answer": "42"
}jsonrollouts.jsonlnew at this step
{"id": "math-0042", "attempt": "A", "completion": "6 × 7 = 42. Answer: 42"}
{"id": "math-0042", "attempt": "B", "completion": "6 + 7 = 13. Answer: 13"}
{"id": "math-0042", "attempt": "C", "completion": "Just a guess. Answer: 42"}
{"id": "math-0042", "attempt": "D", "completion": "6 × 8 = 48. Answer: 48"}pythonreward.pynew at this step
def reward(completion: str, answer: str) -> float:
final = completion.split("Answer:")[-1].strip()
return 1.0 if final == answer else 0.0The whole “judge” is these three lines. That is what makes the reward verifiable.
jsonrollouts.jsonlnew at this step
{"attempt": "A", "completion": "6 × 7 = 42. Answer: 42", "reward": 1.0}
{"attempt": "B", "completion": "6 + 7 = 13. Answer: 13", "reward": 0.0}
{"attempt": "C", "completion": "Just a guess. Answer: 42", "reward": 1.0}
{"attempt": "D", "completion": "6 × 8 = 48. Answer: 48", "reward": 0.0}jsonwhat the update changed
{"final_answer": "42", "p_before": 0.50, "p_after": 0.68}
{"final_answer": "13", "p_before": 0.25, "p_after": 0.17}
{"final_answer": "48", "p_before": 0.25, "p_after": 0.15}Illustrative numbers. The model does not store a table like this; its chances come from its settings, which the update changed.
jsontrain.jsonl · next row
{
"id": "math-0043",
"prompt": "A train travels 60 km per hour for 3 hours. How far does it go? End with 'Answer: <number>'.",
"answer": "180"
}How RLVR works
Choose checkable tasks
Use questions with known answers or code with tests.
Generate attempts
Let the model try answers rather than showing it only the solution.
Run the checker
Turn the test result into a reward.
Update and repeat
Use those rewards to adjust the model, then test on unseen tasks.
Useful for
Math answers, coding tasks, and requirements a reliable program can check.
The catch
Weak checks can be exploited. Passing a few tests does not guarantee correct code, and a correct final answer does not prove every reasoning step was sound.
A little more detail
What does “verifiable” mean?
It means there is a defined way to check the result. It does not mean the model becomes correct about everything.
Can RLVR use GRPO?
Yes. RLVR says where the reward comes from. GRPO says how to use rewards to update the model. The two can be used together.
Who works with RLVR
RL environment builders and AI benchmarks with public evidence that they use RLVR: their own website says so, or two independent sites do. Only confirmed evidence is shown; each page lists the quotes.
RL environments 16
- AfterQuery
- ARIMLABS
- Chakra Labs
Evidence on its own site
“For RLVR tasks the verifier is code, and a judge compares every rollout against its verdict to flag false positives and false negatives, following the DeepSWE methodology.”
chakra.dev - Daytona
Evidence on its own site
“This guide presents a simple, self-contained script that performs reinforcement learning training of Qwen3-1.7B-Base . In particular, we use reinforcement learning with verifiable rewards , with the reward being computed from the test pass rate of model-written functions.”
daytona.io - E2B
Evidence on its own site
“A key part of that pipeline involves reinforcement learning with verifiable rewards , where a large language model (LLM) is trained to solve problems that can be checked for correctness against ground-truth answers. For example, given a simple math problem like “What is 1+1?”, the LLM receives a binary…”
e2b.dev - Halluminate
Evidence on its own site
“We employ RLVR-style unit tests that deterministically evaluate the agent's final state. Complete access to the simulator's underlying data allow us to compute and check ground truth programmatically, eliminating dependence on live-si…”
halluminate.ai - Mechanize
Evidence on its own site
“…environments made sense because AI agents were simply not competent enough to deal with anything resembling the complexity of real-world work settings. However, this is changing, and the new reinforcement learning from verifiable reward (RLVR) paradigm will soon be severely bottlenecked by the lack of a sufficient volume of realistic RL environments. At Mechanize, our immediate goal is to remove this bottleneck and accelerate progress toward a fully au…”
mechanize.work - Prime Intellect
Evidence on its own site
“To train a model capable of this, we use QwQ-32B as our base model and follow Deepseek-R1’s approach of applying GRPO with verifiable rewards from the domains of mathematics and coding. During our preliminary experiments, the following components were crucial to achieve good performance while controlling our model’s thinking budget:”
primeintellect.ai - ReasonCore
Evidence on its own site
“We design and maintain sandboxed RL/RLVR environments that let models learn through interaction with realistic, high-stakes systems. These environments are modeled on enterprise software at its most demanding — operational dashboards, financial platforms, infr…”
reasoncore.ai - Refresh
Evidence on its own site
“An RLVR dataset curated from Gauntlet, delivered in Harbor format for reinforcement learning.”
refresh.dev - Scale AI
Evidence on its own site
“The Future of AI Learning Environments: Verifiable Reward + Multi-Agent Interaction”
scale.com - Snorkel AI
Evidence on its own site
“Learning from Less: Measuring the Effectiveness of RLVR in Low Data and Compute Regimes”
snorkel.ai - Surge AI
Evidence on its own site
“To check whether ComplexConstraints measures something worth optimizing for, we built a companion training set of 1,000 examples, and trained Qwen3-4B using RLVR with a judge model.”
surgehq.ai - Ulam
- Vetto AI
Evidence on its own site
“1. 01RL Environments and AgentsInteractive environments where agents take on open-ended tasks and make real decisions, with verifiers built to reward the behavior that matters. 2. 02Rubrics and VerifiersObjective verifiable reward from what a response gets right and wrong. It's the checkable signal RLHF's preference labeling can't provide. 3. 03Planning & Reasoning TracesTraceable reasoning chains for decisions under uncertainty: maintaining a hy…”
vetto.ai - Vmax
Evidence on its own site
“DescriptionWe introduce PopuLoRA, a population-based asymmetric self-play framework for reinforcement learning with verifiable rewards (RLVR) post-training of LLMs.”
vmax.ai
AI benchmarks 91
- ~1k JavaScript coding tasks in Harbor format
Evidence on its own site
“DescriptionWe introduce PopuLoRA, a population-based asymmetric self-play framework for reinforcement learning with verifiable rewards (RLVR) post-training of LLMs.”
vmax.ai - AdvancedIF - arXiv
Evidence on its own site
“To check whether ComplexConstraints measures something worth optimizing for, we built a companion training set of 1,000 examples, and trained Qwen3-4B using RLVR with a judge model.”
surgehq.ai - Agent-environment papers
Evidence on its own site
“Not Every Rubric Teaches Equally: Policy-Aware Rubric Rewards for RLVR”
labs.scale.com - Agents' Last Exam
Evidence on its own site
“Learning from Less: Measuring the Effectiveness of RLVR in Low Data and Compute Regimes”
snorkel.ai - Antidote and Hemingway-bench leaderboards
Evidence on its own site
“To check whether ComplexConstraints measures something worth optimizing for, we built a companion training set of 1,000 examples, and trained Qwen3-4B using RLVR with a judge model.”
surgehq.ai - App-Bench
- AppBench
- Benchmarking Agents in Insurance Underwriting Environments
Evidence on its own site
“Learning from Less: Measuring the Effectiveness of RLVR in Low Data and Compute Regimes”
snorkel.ai - BigLaw Bench: Research
Evidence on its own site
“Learning from Less: Measuring the Effectiveness of RLVR in Low Data and Compute Regimes”
snorkel.ai - BlenderBench
Evidence on its own site
“An RLVR dataset curated from Gauntlet, delivered in Harbor format for reinforcement learning.”
refresh.dev - BrowserBench (2025-11-22):
Evidence on its own site
“We employ RLVR-style unit tests that deterministically evaluate the agent's final state. Complete access to the simulator's underlying data allow us to compute and check ground truth programmatically, eliminating dependence on live-si…”
halluminate.ai - Cats Confuse Reasoning LLM
Evidence on its own site
“These environments provide what static supervised data cannot: a dynamic sandbox with verifiable rewards at scale. Pass/fail checks, structured scoring, safety constraints, and automated metrics turn RL from a research curiosity into an operational capability for mid-training (inducing new capabilities) and post-training (…”
blog.collinear.ai - Chartography - arXiv:2608.10677
Evidence on its own site
“To check whether ComplexConstraints measures something worth optimizing for, we built a companion training set of 1,000 examples, and trained Qwen3-4B using RLVR with a judge model.”
surgehq.ai - ComplexConstraints and Beyond
Evidence on its own site
“To check whether ComplexConstraints measures something worth optimizing for, we built a companion training set of 1,000 examples, and trained Qwen3-4B using RLVR with a judge model.”
surgehq.ai - computer-1 computer-use harness in Harbor
Evidence on its own site
“An RLVR dataset curated from Gauntlet, delivered in Harbor format for reinforcement learning.”
refresh.dev - Continual Learning Bench
Evidence on its own site
“Learning from Less: Measuring the Effectiveness of RLVR in Low Data and Compute Regimes”
snorkel.ai - CoreCraft
Evidence on its own site
“To check whether ComplexConstraints measures something worth optimizing for, we built a companion training set of 1,000 examples, and trained Qwen3-4B using RLVR with a judge model.”
surgehq.ai - Cross-Benchmark Generalization in Long-Horizon Agents - arXiv
Evidence on its own site
“To check whether ComplexConstraints measures something worth optimizing for, we built a companion training set of 1,000 examples, and trained Qwen3-4B using RLVR with a judge model.”
surgehq.ai - Cross-Benchmark Transfer from RL on Agentic Coding Tasks - arXiv
Evidence on its own site
“To check whether ComplexConstraints measures something worth optimizing for, we built a companion training set of 1,000 examples, and trained Qwen3-4B using RLVR with a judge model.”
surgehq.ai - CWE-bench
Evidence on its own site
“These environments provide what static supervised data cannot: a dynamic sandbox with verifiable rewards at scale. Pass/fail checks, structured scoring, safety constraints, and automated metrics turn RL from a research curiosity into an operational capability for mid-training (inducing new capabilities) and post-training (…”
blog.collinear.ai - DAYJOB (Healthcare/Finance) - arXiv:2610.01306
Evidence on its own site
“To check whether ComplexConstraints measures something worth optimizing for, we built a companion training set of 1,000 examples, and trained Qwen3-4B using RLVR with a judge model.”
surgehq.ai - DealTrace Bench (2026-05/06):
Evidence on its own site
“We employ RLVR-style unit tests that deterministically evaluate the agent's final state. Complete access to the simulator's underlying data allow us to compute and check ground truth programmatically, eliminating dependence on live-si…”
halluminate.ai - Dojo-Bench-Mini public task set on Hugging Face
Evidence on its own site
“For RLVR tasks the verifier is code, and a judge compares every rollout against its verdict to flag false positives and false negatives, following the DeepSWE methodology.”
chakra.dev - DrugDiscoveryBench with Phylo
Evidence on its own site
“Not Every Rubric Teaches Equally: Policy-Aware Rubric Rewards for RLVR”
labs.scale.com - EnterpriseBench CoreCraft - arXiv:2602.16179
Evidence on its own site
“We trained Qwen3-4B (Thinking) with RLVR, on the 1,000-example companion set with a 90:10 train/test split. An LLM judge grades each criterion as a binary pass/fail per rollout, and the reward is the fraction of criteria satisfied.”
surgehq.ai - FinanceQA
- FinQA / SnorkelFinance environment and benchmark
Evidence on its own site
“Learning from Less: Measuring the Effectiveness of RLVR in Low Data and Compute Regimes”
snorkel.ai - Gauntlet 4K RLVR dataset, access by request
Evidence on its own site
“An RLVR dataset curated from Gauntlet, delivered in Harbor format for reinforcement learning.”
refresh.dev - GBA Eval
Evidence on its own site
“…environments made sense because AI agents were simply not competent enough to deal with anything resembling the complexity of real-world work settings. However, this is changing, and the new reinforcement learning from verifiable reward (RLVR) paradigm will soon be severely bottlenecked by the lack of a sufficient volume of realistic RL environments. At Mechanize, our immediate goal is to remove this bottleneck and accelerate progress toward a fully au…”
mechanize.work - gba-eval-attempts and per-model gba-eval-attempt-* repos
Evidence on its own site
“…environments made sense because AI agents were simply not competent enough to deal with anything resembling the complexity of real-world work settings. However, this is changing, and the new reinforcement learning from verifiable reward (RLVR) paradigm will soon be severely bottlenecked by the lack of a sufficient volume of realistic RL environments. At Mechanize, our immediate goal is to remove this bottleneck and accelerate progress toward a fully au…”
mechanize.work - GDP.pdf - arXiv:2607.11192
Evidence on its own site
“We trained Qwen3-4B (Thinking) with RLVR, on the 1,000-example companion set with a 90:10 train/test split. An LLM judge grades each criterion as a binary pass/fail per rollout, and the reward is the fraction of criteria satisfied.”
surgehq.ai - GDP.xlsx
Evidence on its own site
“We trained Qwen3-4B (Thinking) with RLVR, on the 1,000-example companion set with a 90:10 train/test split. An LLM judge grades each criterion as a binary pass/fail per rollout, and the reward is the fraction of criteria satisfied.”
surgehq.ai - General Agent synthetic environment
Evidence on its own site
“To train a model capable of this, we use QwQ-32B as our base model and follow Deepseek-R1’s approach of applying GRPO with verifiable rewards from the domains of mathematics and coding. During our preliminary experiments, the following components were crucial to achieve good performance while controlling our model’s thinking budget:”
primeintellect.ai - GPU Glossary
Evidence on its own site
“Reinforcement learning with verifiable rewards (RLVR) depends on reliable signals for mathematical reasoning, code correctness, and step-level progress. This guide compares six leading open-source verifier models and the Modal infrastructure teams use to train, serv…”
modal.com - HANDBOOK.md - arXiv:2607.25398
Evidence on its own site
“We trained Qwen3-4B (Thinking) with RLVR, on the 1,000-example companion set with a 90:10 train/test split. An LLM judge grades each criterion as a binary pass/fail per rollout, and the reward is the fraction of criteria satisfied.”
surgehq.ai - HANDBOOK.md Agents
Evidence on its own site
“We trained Qwen3-4B (Thinking) with RLVR, on the 1,000-example companion set with a 90:10 train/test split. An LLM judge grades each criterion as a binary pass/fail per rollout, and the reward is the fraction of criteria satisfied.”
surgehq.ai - HiL-Bench
Evidence on its own site
“Not Every Rubric Teaches Equally: Policy-Aware Rubric Rewards for RLVR”
labs.scale.com - Human Evaluation
Evidence on its own site
“We trained Qwen3-4B (Thinking) with RLVR, on the 1,000-example companion set with a 90:10 train/test split. An LLM judge grades each criterion as a binary pass/fail per rollout, and the reward is the fraction of criteria satisfied.”
surgehq.ai - Humanity's Last Exam
Evidence on its own site
“Not Every Rubric Teaches Equally: Policy-Aware Rubric Rewards for RLVR”
labs.scale.com - IDE-Bench
- INTELLECT-2
Evidence on its own site
“To train a model capable of this, we use QwQ-32B as our base model and follow Deepseek-R1’s approach of applying GRPO with verifiable rewards from the domains of mathematics and coding. During our preliminary experiments, the following components were crucial to achieve good performance while controlling our model’s thinking budget:”
primeintellect.ai - INTELLECT-3 Technical Report
Evidence on its own site
“To train a model capable of this, we use QwQ-32B as our base model and follow Deepseek-R1’s approach of applying GRPO with verifiable rewards from the domains of mathematics and coding. During our preliminary experiments, the following components were crucial to achieve good performance while controlling our model’s thinking budget:”
primeintellect.ai - Kubernetes Is the Wrong Primitive for Diverse Workloads
Evidence on its own site
“This guide presents a simple, self-contained script that performs reinforcement learning training of Qwen3-1.7B-Base . In particular, we use reinforcement learning with verifiable rewards , with the reward being computed from the test pass rate of model-written functions.”
daytona.io - Learning from Less: RLVR in Low Data Regimes
Evidence on its own site
“Learning from Less: Measuring the Effectiveness of RLVR in Low Data and Compute Regimes”
snorkel.ai - Legora BAR
- LibraryDesignBench
Evidence on its own site
“Learning from Less: Measuring the Effectiveness of RLVR in Low Data and Compute Regimes”
snorkel.ai - Market-Bench
- Math RL Tasks
- MCP Atlas
Evidence on its own site
“Not Every Rubric Teaches Equally: Policy-Aware Rubric Rewards for RLVR”
labs.scale.com - Measuring Autonomous AI Research
Evidence on its own site
“To train a model capable of this, we use QwQ-32B as our base model and follow Deepseek-R1’s approach of applying GRPO with verifiable rewards from the domains of mathematics and coding. During our preliminary experiments, the following components were crucial to achieve good performance while controlling our model’s thinking budget:”
primeintellect.ai - OpenReward
Evidence on its own site
“…ting LLM tool use with reinforcement learning, generating synthetic scenarios, integrated functions, complexity scaling, and localized deployments to create stable, high-quality training environments. It also includes a verifiable reward mechanism that quantifies both precision of tool use and completeness of task execution, enabling feedback-driven RL training and measurable evaluation of tool-use performance.”
openreward.ai - OSWorld 2.0
Evidence on its own site
“Learning from Less: Measuring the Effectiveness of RLVR in Low Data and Compute Regimes”
snorkel.ai - OSWorld Windows tasks on Daytona Windows sandboxes
Evidence on its own site
“This guide presents a simple, self-contained script that performs reinforcement learning training of Qwen3-1.7B-Base . In particular, we use reinforcement learning with verifiable rewards , with the reward being computed from the test pass rate of model-written functions.”
daytona.io - PopuLoRA
Evidence on its own site
“DescriptionWe introduce PopuLoRA, a population-based asymmetric self-play framework for reinforcement learning with verifiable rewards (RLVR) post-training of LLMs.”
vmax.ai - Post-Training on Office Work Improves Software Engineering - arXiv
Evidence on its own site
“We trained Qwen3-4B (Thinking) with RLVR, on the 1,000-example companion set with a 90:10 train/test split. An LLM judge grades each criterion as a binary pass/fail per rollout, and the reward is the fraction of criteria satisfied.”
surgehq.ai - PRBench
Evidence on its own site
“Not Every Rubric Teaches Equally: Policy-Aware Rubric Rewards for RLVR”
labs.scale.com - PROPEL
Evidence on its own site
“DescriptionWe introduce PopuLoRA, a population-based asymmetric self-play framework for reinforcement learning with verifiable rewards (RLVR) post-training of LLMs.”
vmax.ai - Quail open-source AI-SQL inference engine with CMU Full Stack Data Lab
Evidence on its own site
“Reinforcement learning with verifiable rewards (RLVR) depends on reliable signals for mathematical reasoning, code correctness, and step-level progress. This guide compares six leading open-source verifier models and the Modal infrastructure teams use to train, serv…”
modal.com - READY or Not enterprise agent deployment benchmarks
Evidence on its own site
“Not Every Rubric Teaches Equally: Policy-Aware Rubric Rewards for RLVR”
labs.scale.com - Remote Labor Index
Evidence on its own site
“Not Every Rubric Teaches Equally: Policy-Aware Rubric Rewards for RLVR”
labs.scale.com - Riemann-Bench - arXiv:2604.06802
Evidence on its own site
“We trained Qwen3-4B (Thinking) with RLVR, on the 1,000-example companion set with a 90:10 train/test split. An LLM judge grades each criterion as a binary pass/fail per rollout, and the reward is the fraction of criteria satisfied.”
surgehq.ai - RIFT / Hitchhiker's Guide to Rubric Quality
Evidence on its own site
“Learning from Less: Measuring the Effectiveness of RLVR in Low Data and Compute Regimes”
snorkel.ai - RSI Bench
Evidence on its own site
“Not Every Rubric Teaches Equally: Policy-Aware Rubric Rewards for RLVR”
labs.scale.com - Rubric-based RL papers
Evidence on its own site
“Not Every Rubric Teaches Equally: Policy-Aware Rubric Rewards for RLVR”
labs.scale.com - Scale Labs (launched 2026-03-09
Evidence on its own site
“Not Every Rubric Teaches Equally: Policy-Aware Rubric Rewards for RLVR”
labs.scale.com - Scaling Agentic RL: 365k+ SWE/terminal/search tasks
Evidence on its own site
“To train a model capable of this, we use QwQ-32B as our base model and follow Deepseek-R1’s approach of applying GRPO with verifiable rewards from the domains of mathematics and coding. During our preliminary experiments, the following components were crucial to achieve good performance while controlling our model’s thinking budget:”
primeintellect.ai - Searching Over Sandbox States for Coding Agents
Evidence on its own site
“This guide presents a simple, self-contained script that performs reinforcement learning training of Qwen3-1.7B-Base . In particular, we use reinforcement learning with verifiable rewards , with the reward being computed from the test pass rate of model-written functions.”
daytona.io - Senior SWE-bench
Evidence on its own site
“Learning from Less: Measuring the Effectiveness of RLVR in Low Data and Compute Regimes”
snorkel.ai - Shrinking the Generation-Verification Gap with Weak Verifiers
Evidence on its own site
“Learning from Less: Measuring the Effectiveness of RLVR in Low Data and Compute Regimes”
snorkel.ai - Simulated-to-Real Transfer for Web Agents (2026-07-27):
Evidence on its own site
“We employ RLVR-style unit tests that deterministically evaluate the agent's final state. Complete access to the simulator's underlying data allow us to compute and check ground truth programmatically, eliminating dependence on live-si…”
halluminate.ai - SlopCodeBench
Evidence on its own site
“Learning from Less: Measuring the Effectiveness of RLVR in Low Data and Compute Regimes”
snorkel.ai - Snorkel: Rapid Training Data Creation with Weak Supervision
Evidence on its own site
“Learning from Less: Measuring the Effectiveness of RLVR in Low Data and Compute Regimes”
snorkel.ai - SpreadsheetBench 2
- SWE-bench Pro
Evidence on its own site
“Not Every Rubric Teaches Equally: Policy-Aware Rubric Rewards for RLVR”
labs.scale.com - SWE-Bench Pro - Private Dataset
Evidence on its own site
“Not Every Rubric Teaches Equally: Policy-Aware Rubric Rewards for RLVR”
labs.scale.com - SWE-Bench Pro v1 (2025-09-19) and v2
Evidence on its own site
“Not Every Rubric Teaches Equally: Policy-Aware Rubric Rewards for RLVR”
labs.scale.com - Systematic Reward Hacking study
Evidence on its own site
“To train a model capable of this, we use QwQ-32B as our base model and follow Deepseek-R1’s approach of applying GRPO with verifiable rewards from the domains of mathematics and coding. During our preliminary experiments, the following components were crucial to achieve good performance while controlling our model’s thinking budget:”
primeintellect.ai - Terminal-Bench
Evidence on its own site
“Learning from Less: Measuring the Effectiveness of RLVR in Low Data and Compute Regimes”
snorkel.ai - The Hierarchy of Agentic Capabilities - arXiv:2601.09032
Evidence on its own site
“We trained Qwen3-4B (Thinking) with RLVR, on the 1,000-example companion set with a 90:10 train/test split. An LLM judge grades each criterion as a binary pass/fail per rollout, and the reward is the fraction of criteria satisfied.”
surgehq.ai - The Valley of Code Reasoning
Evidence on its own site
“These environments provide what static supervised data cannot: a dynamic sandbox with verifiable rewards at scale. Pass/fail checks, structured scoring, safety constraints, and automated metrics turn RL from a research curiosity into an operational capability for mid-training (inducing new capabilities) and post-training (…”
blog.collinear.ai - TraitBasis / Impatient Users Confuse AI Agents
Evidence on its own site
“These environments provide what static supervised data cannot: a dynamic sandbox with verifiable rewards at scale. Pass/fail checks, structured scoring, safety constraints, and automated metrics turn RL from a research curiosity into an operational capability for mid-training (inducing new capabilities) and post-training (…”
blog.collinear.ai - True Agents Model the World / ECHO
Evidence on its own site
“To train a model capable of this, we use QwQ-32B as our base model and follow Deepseek-R1’s approach of applying GRPO with verifiable rewards from the domains of mathematics and coding. During our preliminary experiments, the following components were crucial to achieve good performance while controlling our model’s thinking budget:”
primeintellect.ai - Tuesday Work Index leaderboard
Evidence on its own site
“We trained Qwen3-4B (Thinking) with RLVR, on the 1,000-example companion set with a 90:10 train/test split. An LLM judge grades each criterion as a binary pass/fail per rollout, and the reward is the fraction of criteria satisfied.”
surgehq.ai - Universal offline sandbox escape disclosure
Evidence on its own site
“To train a model capable of this, we use QwQ-32B as our base model and follow Deepseek-R1’s approach of applying GRPO with verifiable rewards from the domains of mathematics and coding. During our preliminary experiments, the following components were crucial to achieve good performance while controlling our model’s thinking budget:”
primeintellect.ai - unix-ctf
Evidence on its own site
“…procedurally generated CTF environments, unix-ctf creates a focused training and evaluation surface for Unix competence. The generator produces a reusable library at high yield, the multi-flag environments provide dense verifiable reward, and GRPO training on the surface improves held-out Unix-style recovery tasks.”
vmax.ai - VERITAS
Evidence on its own site
“These environments provide what static supervised data cannot: a dynamic sandbox with verifiable rewards at scale. Pass/fail checks, structured scoring, safety constraints, and automated metrics turn RL from a research curiosity into an operational capability for mid-training (inducing new capabilities) and post-training (…”
blog.collinear.ai - Web Bench (2025-05-30; ~5
Evidence on its own site
“We employ RLVR-style unit tests that deterministically evaluate the agent's final state. Complete access to the simulator's underlying data allow us to compute and check ground truth programmatically, eliminating dependence on live-si…”
halluminate.ai - WebBench
Evidence on its own site
“We employ RLVR-style unit tests that deterministically evaluate the agent's final state. Complete access to the simulator's underlying data allow us to compute and check ground truth programmatically, eliminating dependence on live-si…”
halluminate.ai - Westworld Finance Diligence Bench (2026-08-11):
Evidence on its own site
“We employ RLVR-style unit tests that deterministically evaluate the agent's final state. Complete access to the simulator's underlying data allow us to compute and check ground truth programmatically, eliminating dependence on live-si…”
halluminate.ai - Westworld v1 simulated-web environments (2025-11-23):
Evidence on its own site
“We employ RLVR-style unit tests that deterministically evaluate the agent's final state. Complete access to the simulator's underlying data allow us to compute and check ground truth programmatically, eliminating dependence on live-si…”
halluminate.ai - YC-Bench
Evidence on its own site
“These environments provide what static supervised data cannot: a dynamic sandbox with verifiable rewards at scale. Pass/fail checks, structured scoring, safety constraints, and automated metrics turn RL from a research curiosity into an operational capability for mid-training (inducing new capabilities) and post-training (…”
blog.collinear.ai
Read the research
These guides simplify the methods. The papers describe the full training recipes. Examples on this page are illustrative, not experimental results.