Reinforced Labs: research and learn https://reinforcedlabs.ai/learn Research notes and guides on reinforcement learning, RL for language models, AI safety and AI evals. en Sun, 01 Nov 2026 12:00:00 GMT What is reinforcement learning? A guide from gridworlds to language modelshttps://reinforcedlabs.ai/learn/what-is-reinforcement-learninghttps://reinforcedlabs.ai/learn/what-is-reinforcement-learningSun, 01 Nov 2026 12:00:00 GMTReinforcement learning explained with a gridworld you can play with: agents, rewards, policies, values, planning vs Q-learning, algorithm types, pitfalls.Reinforcement learning What is AI safety? Risks, research areas and how evaluation fitshttps://reinforcedlabs.ai/learn/what-is-ai-safetyhttps://reinforcedlabs.ai/learn/what-is-ai-safetySun, 01 Nov 2026 12:00:00 GMTWhat AI safety is: misuse, accident and structural risks, the technical sub-areas from alignment to oversight, and how evaluation turns them into evidence.AI safety What are AI evals? A complete guide to LLM evals: types, components and how to build onehttps://reinforcedlabs.ai/learn/what-are-ai-evalshttps://reinforcedlabs.ai/learn/what-are-ai-evalsSun, 01 Nov 2026 12:00:00 GMTAI evals explained: capability, safety and product evals, the four components, sources of error, offline vs online evaluation, and how to build one.Evaluations Policy gradient baselines and the advantage function: why they reduce variance without biashttps://reinforcedlabs.ai/learn/variance-reduction-baselineshttps://reinforcedlabs.ai/learn/variance-reduction-baselinesSun, 01 Nov 2026 12:00:00 GMTWhy subtracting a baseline leaves the policy gradient unbiased, what the advantage function is, the optimal baseline, and measured variance reductions.Reinforcement learning Value iteration vs policy iteration: two ways to solve an MDPhttps://reinforcedlabs.ai/learn/value-iteration-policy-iterationhttps://reinforcedlabs.ai/learn/value-iteration-policy-iterationSun, 01 Nov 2026 12:00:00 GMTValue iteration and policy iteration side by side: equations, a worked gridworld, stopping rules, modified policy iteration and timed runs on large grids.Reinforcement learning PPO in reinforcement learning, from first principles: trust regions, natural gradients and clippinghttps://reinforcedlabs.ai/learn/trpo-ppo-explainedhttps://reinforcedlabs.ai/learn/trpo-ppo-explainedSun, 01 Nov 2026 12:00:00 GMTPPO from first principles: the surrogate objective, TRPO's KL trust region, the natural gradient, the clipped objective, and measured KL per update.Reinforcement learning Statistical significance for LLM evaluation: paired tests, bootstrap, multiple comparisons and powerhttps://reinforcedlabs.ai/learn/statistical-significance-for-evalshttps://reinforcedlabs.ai/learn/statistical-significance-for-evalsSun, 01 Nov 2026 12:00:00 GMTWhen is a gap between two eval scores real? Paired tests, bootstrap, repeated sampling, multiple-comparison corrections and power, with simulations.Evaluations Specification gaming and Goodhart's law in AIhttps://reinforcedlabs.ai/learn/specification-gaminghttps://reinforcedlabs.ai/learn/specification-gamingSun, 01 Nov 2026 12:00:00 GMTWhat specification gaming is, Goodhart's four variants, a formal toy model of when proxies collapse under optimisation, and published examples.AI safety Scalable oversight: debate, amplification, weak-to-strong generalisation and sandwichinghttps://reinforcedlabs.ai/learn/scalable-oversighthttps://reinforcedlabs.ai/learn/scalable-oversightSun, 01 Nov 2026 12:00:00 GMTHow to supervise AI on tasks the supervisor cannot check: debate, amplification, recursive reward modelling, critique, weak-to-strong and sandwiching.AI safety Safe reinforcement learning: constrained MDPs, Lagrangian methods and shieldinghttps://reinforcedlabs.ai/learn/safe-reinforcement-learninghttps://reinforcedlabs.ai/learn/safe-reinforcement-learningSun, 01 Nov 2026 12:00:00 GMTSafe RL explained: constrained MDPs, Lagrangian primal-dual training and why it oscillates, constraints during training, shielding and safe exploration.AI safety RLHF explained: how reinforcement learning from human feedback works, and where it failshttps://reinforcedlabs.ai/learn/rlhf-explainedhttps://reinforcedlabs.ai/learn/rlhf-explainedSun, 01 Nov 2026 12:00:00 GMTHow RLHF works end to end: SFT, preference data, Bradley-Terry reward models, PPO with a KL penalty, a toy pipeline, failure modes and safety limits.RL for language models RLAIF and Constitutional AI: reinforcement learning from AI feedbackhttps://reinforcedlabs.ai/learn/rlaif-constitutional-aihttps://reinforcedlabs.ai/learn/rlaif-constitutional-aiSun, 01 Nov 2026 12:00:00 GMTHow RLAIF and Constitutional AI replace human labels with model judgments: critique and revision, AI preference labels, evidence vs RLHF, failures.RL for language models Reinforcement learning with verifiable rewards (RLVR), explainedhttps://reinforcedlabs.ai/learn/rl-with-verifiable-rewardshttps://reinforcedlabs.ai/learn/rl-with-verifiable-rewardsSun, 01 Nov 2026 12:00:00 GMTHow RLVR trains models with programmatic verifiers: group-relative advantages, why pass rate controls the learning signal, and what verifier errors do.RL for language models RL environments for LLM agents: state, tools, reward and sandboxinghttps://reinforcedlabs.ai/learn/rl-environments-for-agentshttps://reinforcedlabs.ai/learn/rl-environments-for-agentsSun, 01 Nov 2026 12:00:00 GMTHow to design RL environments for LLM agents: tasks, state, tool calls, rewards, sandboxing, reward leakage, and why eval and training environments differ.RL for language models Reward shaping: why potential-based shaping is safe and other bonuses can break your agenthttps://reinforcedlabs.ai/learn/reward-shapinghttps://reinforcedlabs.ai/learn/reward-shapingSun, 01 Nov 2026 12:00:00 GMTWhat reward shaping is, why potential-based shaping (Ng et al. 1999) keeps optimal policies, how other bonuses create loops, and its effect on learning.Reinforcement learning Reward models: how they are trained, how to evaluate them, and why accuracy is not enoughhttps://reinforcedlabs.ai/learn/reward-modelshttps://reinforcedlabs.ai/learn/reward-modelsSun, 01 Nov 2026 12:00:00 GMTWhat a reward model is, how Bradley-Terry training works, how to evaluate reward models, and why pairwise accuracy can miss what matters downstream.RL for language models Reward hacking: definition, taxonomy, examples and mitigationhttps://reinforcedlabs.ai/learn/reward-hackinghttps://reinforcedlabs.ai/learn/reward-hackingSun, 01 Nov 2026 12:00:00 GMTWhat reward hacking is, a taxonomy of how rewards get exploited, documented examples from RL and LLM training, and how to detect and prevent it.RL for language models Reinforcement learning for LLMs: how RL trains language models, from RLHF to RLVRhttps://reinforcedlabs.ai/learn/reinforcement-learning-for-llmshttps://reinforcedlabs.ai/learn/reinforcement-learning-for-llmsSun, 01 Nov 2026 12:00:00 GMTHow RL trains LLMs: the model as a policy, tokens as actions, what RL adds over SFT, RLHF to RLVR, and why reward hacking is a safety problem.RL for language models The REINFORCE algorithm: derivation, pseudocode, numpy code and its variance problemhttps://reinforcedlabs.ai/learn/reinforce-algorithmhttps://reinforcedlabs.ai/learn/reinforce-algorithmSun, 01 Nov 2026 12:00:00 GMTWilliams's REINFORCE explained: the update, pseudocode, runnable numpy code, and measured gradient noise that shows why plain REINFORCE learns slowly.Reinforcement learning Q-learning explained: the update rule, worked examples and DQNhttps://reinforcedlabs.ai/learn/q-learninghttps://reinforcedlabs.ai/learn/q-learningSun, 01 Nov 2026 12:00:00 GMTQ-learning from first principles: the update rule worked by hand, a live agent, seeded results on step size, SARSA on the cliff, double Q-learning and DQN.Reinforcement learning The policy gradient theorem: a full derivation, checked numericallyhttps://reinforcedlabs.ai/learn/policy-gradient-theoremhttps://reinforcedlabs.ai/learn/policy-gradient-theoremSun, 01 Nov 2026 12:00:00 GMTDerive the policy gradient theorem from the log-derivative trick, see why dynamics drop out, and check every step on a two-state MDP you can solve by hand.Reinforcement learning Offline reinforcement learning: learning from logged data without falling off the cliffhttps://reinforcedlabs.ai/learn/offline-reinforcement-learninghttps://reinforcedlabs.ai/learn/offline-reinforcement-learningSun, 01 Nov 2026 12:00:00 GMTWhat offline RL is, why distribution shift makes Q-learning overrate unseen actions, and how CQL, IQL and behaviour regularisation fix it. Live demo.Reinforcement learning Multi-armed bandits: epsilon-greedy, UCB and Thompson sampling, explained and measuredhttps://reinforcedlabs.ai/learn/multi-armed-banditshttps://reinforcedlabs.ai/learn/multi-armed-banditsSun, 01 Nov 2026 12:00:00 GMTWhat a multi-armed bandit is, how regret works, and how epsilon-greedy, UCB and Thompson sampling compare in seeded simulations, including vs A/B tests.Reinforcement learning Markov decision process (MDP): definition, example and how to solve onehttps://reinforcedlabs.ai/learn/markov-decision-processhttps://reinforcedlabs.ai/learn/markov-decision-processSun, 01 Nov 2026 12:00:00 GMTWhat a Markov decision process is, its five components, the Markov property and discounting, with an interactive gridworld and seeded simulations.Reinforcement learning LLM benchmarks explained: what they measure, saturation, contamination and how to read leaderboardshttps://reinforcedlabs.ai/learn/llm-benchmarkshttps://reinforcedlabs.ai/learn/llm-benchmarksSun, 01 Nov 2026 12:00:00 GMTWhat MMLU, GPQA, GSM8K, HumanEval, SWE-bench, IFEval and Chatbot Arena measure, why benchmarks saturate and leak, and how to read a leaderboard.Evaluations KL penalty in RLHF: KL-regularised RL, the optimal policy and how to choose betahttps://reinforcedlabs.ai/learn/kl-regularized-rlhttps://reinforcedlabs.ai/learn/kl-regularized-rlSun, 01 Nov 2026 12:00:00 GMTWhy RLHF penalises KL to a reference model, the closed-form optimal policy, how beta trades reward for drift, k1/k2/k3 estimators and adaptive KL control.RL for language models LLM jailbreaks explained: why they work, attack families and how success is measuredhttps://reinforcedlabs.ai/learn/jailbreaks-explainedhttps://reinforcedlabs.ai/learn/jailbreaks-explainedSun, 01 Nov 2026 12:00:00 GMTWhat an LLM jailbreak is, why safety training fails, the main attack families at a conceptual level, and why attack success rate needs a stated budget.AI safety Inverse reinforcement learning: inferring the reward behind the behaviourhttps://reinforcedlabs.ai/learn/inverse-reinforcement-learninghttps://reinforcedlabs.ai/learn/inverse-reinforcement-learningSun, 01 Nov 2026 12:00:00 GMTWhat inverse RL is, why many rewards explain the same demonstrations, how max-entropy IRL works, and how it leads to learning from preferences.Reinforcement learning How to evaluate reinforcement learning agents: seeds, IQM and confidence intervalshttps://reinforcedlabs.ai/learn/evaluating-rl-agentshttps://reinforcedlabs.ai/learn/evaluating-rl-agentsSun, 01 Nov 2026 12:00:00 GMTSeeds, normalised scores, IQM, stratified bootstrap intervals and performance profiles, tested on 1,500 real Q-learning runs you can resample.Evaluations LLM evaluation metrics: accuracy, F1, exact match, pass@k, ECE, win rate, BLEU and ROUGEhttps://reinforcedlabs.ai/learn/eval-metricshttps://reinforcedlabs.ai/learn/eval-metricsSun, 01 Nov 2026 12:00:00 GMTAccuracy, F1, exact match, pass@k, calibration error, win rate, BLEU and ROUGE: definitions, where each misleads, and which to use for which task.Evaluations Dangerous capability evaluations: what they measure, elicitation, thresholds and uncertaintyhttps://reinforcedlabs.ai/learn/dangerous-capability-evaluationshttps://reinforcedlabs.ai/learn/dangerous-capability-evaluationsSun, 01 Nov 2026 12:00:00 GMTWhat dangerous capability evaluations measure, why elicitation and sandbagging dominate the error, how thresholds work, and why more tasks cannot fix bias.AI safety The Bellman equation: derivation, intuition and a worked examplehttps://reinforcedlabs.ai/learn/bellman-equationhttps://reinforcedlabs.ai/learn/bellman-equationSun, 01 Nov 2026 12:00:00 GMTBellman expectation and optimality equations derived step by step, a backup worked with real numbers, matrix form, contraction and an interactive demo.Reinforcement learning AI safety benchmarks: what they measure, over-refusal, validity and how to choosehttps://reinforcedlabs.ai/learn/ai-safety-benchmarkshttps://reinforcedlabs.ai/learn/ai-safety-benchmarksSun, 01 Nov 2026 12:00:00 GMTWhat the main public AI safety benchmarks measure, why refusal needs an over-refusal pair, how capability leaks into safety scores, and how to choose.AI safety AI red teaming vs evals: goals, methods, metrics and how they work togetherhttps://reinforcedlabs.ai/learn/ai-red-teaming-vs-evalshttps://reinforcedlabs.ai/learn/ai-red-teaming-vs-evalsSun, 01 Nov 2026 12:00:00 GMTEvals measure failure rates on fixed inputs; red teams hunt for failures adaptively. A side-by-side comparison, a simulation, and how to combine them.Evaluations AI alignment explained: outer vs inner alignment, methods and open problemshttps://reinforcedlabs.ai/learn/ai-alignmenthttps://reinforcedlabs.ai/learn/ai-alignmentSun, 01 Nov 2026 12:00:00 GMTWhat AI alignment means, outer vs inner alignment with a goal misgeneralisation toy you can run, the methods used today and the problems still open.AI safety Actor-critic methods and generalized advantage estimation (GAE), with the λ trade-off measuredhttps://reinforcedlabs.ai/learn/actor-critic-gaehttps://reinforcedlabs.ai/learn/actor-critic-gaeSun, 01 Nov 2026 12:00:00 GMTHow actor-critic methods use a critic and the TD error, how GAE blends n-step advantages with lambda, and measured bias and variance for every lambda.Reinforcement learning AI compliance testing: EU AI Act, NIST AI RMF and ISO/IEC 42001https://reinforcedlabs.ai/learn/ai-compliance-testinghttps://reinforcedlabs.ai/learn/ai-compliance-testingTue, 27 Oct 2026 12:00:00 GMTWhat each framework actually asks you to test, mapped to concrete evaluations and evidence. LLM-as-a-judge: how it works, where it fails, how to validate ithttps://reinforcedlabs.ai/learn/llm-as-a-judgehttps://reinforcedlabs.ai/learn/llm-as-a-judgeSat, 24 Oct 2026 12:00:00 GMTJudge designs, documented biases, and a validation protocol with agreement metrics. How many test cases does an LLM eval need?https://reinforcedlabs.ai/research/eval-sample-sizehttps://reinforcedlabs.ai/research/eval-sample-sizeWed, 21 Oct 2026 12:00:00 GMTWilson intervals, power analysis and a Monte Carlo study of paired vs. unpaired tests. Detecting a 2-point regression takes thousands of cases, not hundreds.Evaluation science EU AI Act testing requirements and the 2026–2028 timelinehttps://reinforcedlabs.ai/learn/eu-ai-act-testing-requirementshttps://reinforcedlabs.ai/learn/eu-ai-act-testing-requirementsTue, 20 Oct 2026 12:00:00 GMTArticles 9, 15 and 55, the Digital Omnibus delay, and what evidence a test programme should produce. LLM-as-a-judge vs. human evaluationhttps://reinforcedlabs.ai/compare/llm-as-a-judge-vs-human-evaluationhttps://reinforcedlabs.ai/compare/llm-as-a-judge-vs-human-evaluationSat, 17 Oct 2026 12:00:00 GMTCost, agreement, latency and failure modes compared, and a hybrid design that uses each where it is strongest. AI safety and evaluation for financial serviceshttps://reinforcedlabs.ai/industries/financial-serviceshttps://reinforcedlabs.ai/industries/financial-servicesFri, 16 Oct 2026 12:00:00 GMTModel risk management (SR 11-7), credit and fraud use cases, and agent controls for payments. Your LLM judge is a diagnostic test: correcting measured failure rateshttps://reinforcedlabs.ai/research/llm-judge-errorhttps://reinforcedlabs.ai/research/llm-judge-errorWed, 14 Oct 2026 12:00:00 GMTSensitivity, specificity and the Rogan–Gladen correction. At a 0.5% true failure rate, a 97%-accurate judge reports 3.4%.Evaluation science Evaluating agents against adaptive, multi-turn attackshttps://reinforcedlabs.ai/research/adaptive-attack-evalshttps://reinforcedlabs.ai/research/adaptive-attack-evalsTue, 13 Oct 2026 12:00:00 GMTWhy static attack suites overstate defences, how success grows over turns, why attack budgets must be reported, and a protocol that avoids overfitting.Red teaming and security NIST AI RMF Measure for generative AI: an evaluation checklisthttps://reinforcedlabs.ai/learn/nist-ai-rmf-measurehttps://reinforcedlabs.ai/learn/nist-ai-rmf-measureTue, 13 Oct 2026 12:00:00 GMTTurning the Measure function and the NIST AI 600-1 profile into tests you can run. Indirect prompt injection, explainedhttps://reinforcedlabs.ai/learn/indirect-prompt-injectionhttps://reinforcedlabs.ai/learn/indirect-prompt-injectionSat, 10 Oct 2026 12:00:00 GMTHow instructions hidden in documents, web pages and tool output hijack agents, with real research and the controls that hold. LLM evaluation plan templatehttps://reinforcedlabs.ai/templates/llm-eval-planhttps://reinforcedlabs.ai/templates/llm-eval-planFri, 09 Oct 2026 12:00:00 GMTA copy-ready plan: scope, threat model, categories, sample sizes, graders, gates and owners. pass@k vs. pass^k: measuring agent reliabilityhttps://reinforcedlabs.ai/research/pass-at-k-vs-pass-hat-khttps://reinforcedlabs.ai/research/pass-at-k-vs-pass-hat-kWed, 07 Oct 2026 12:00:00 GMTThree task populations with identical pass@1 = 0.60 have pass^8 between 1.7% and 28.8%. How to estimate pass^k and how many trials you need.Evaluation science Evaluating clinical and healthcare LLM applicationshttps://reinforcedlabs.ai/industries/healthcarehttps://reinforcedlabs.ai/industries/healthcareTue, 06 Oct 2026 12:00:00 GMTClinical harm taxonomies, expert graders, over-refusal and documentation for regulated deployments. Prompt injection vs. jailbreak: what's the difference?https://reinforcedlabs.ai/compare/prompt-injection-vs-jailbreakhttps://reinforcedlabs.ai/compare/prompt-injection-vs-jailbreakSat, 03 Oct 2026 12:00:00 GMTDifferent attacker, different target, different defences. A precise comparison with examples. AI red-team report templatehttps://reinforcedlabs.ai/templates/ai-red-team-reporthttps://reinforcedlabs.ai/templates/ai-red-team-reportFri, 02 Oct 2026 12:00:00 GMTFinding format, severity model, reproduction rate and retest log, ready to copy. A 2% attack success rate is not low: retry economics of prompt injectionhttps://reinforcedlabs.ai/research/prompt-injection-retry-riskhttps://reinforcedlabs.ai/research/prompt-injection-retry-riskWed, 30 Sep 2026 12:00:00 GMTCumulative compromise probability, and why correlated defence layers deliver 3–30× less protection than the independence assumption predicts.Red teaming and security OWASP Top 10 for LLM Applications (2025), explained for engineershttps://reinforcedlabs.ai/learn/owasp-top-10-llmhttps://reinforcedlabs.ai/learn/owasp-top-10-llmSat, 26 Sep 2026 12:00:00 GMTAll ten risks, what each looks like in an agent, and the test and control that addresses it. Guardrail thresholds under low base rateshttps://reinforcedlabs.ai/research/guardrail-thresholdshttps://reinforcedlabs.ai/research/guardrail-thresholdsWed, 23 Sep 2026 12:00:00 GMTWith a 0.1% attack rate, a classifier with AUC 0.96 tuned to 95% recall has 0.5% precision. Cost-optimal thresholds, worked through.Guardrails and monitoring MCP security: tool poisoning, rug pulls and how to defendhttps://reinforcedlabs.ai/learn/mcp-securityhttps://reinforcedlabs.ai/learn/mcp-securitySat, 19 Sep 2026 12:00:00 GMTThe Model Context Protocol attack surface and a hardening checklist for teams connecting agents to MCP servers. The latency budget of AI guardrails: sequential, parallel and hedgedhttps://reinforcedlabs.ai/research/guardrail-latencyhttps://reinforcedlabs.ai/research/guardrail-latencyWed, 16 Sep 2026 12:00:00 GMTOrder statistics of parallel checks. Hedging at the single-check p95 cuts eight-check p99 from 49 ms to 35 ms for 5% extra load.Guardrails and monitoring Evals vs. guardrails vs. red teaminghttps://reinforcedlabs.ai/compare/evals-vs-guardrailshttps://reinforcedlabs.ai/compare/evals-vs-guardrailsTue, 15 Sep 2026 12:00:00 GMTComparison. Building training and evaluation data in regulated domainshttps://reinforcedlabs.ai/research/sensitive-domain-data-pipelineshttps://reinforcedlabs.ai/research/sensitive-domain-data-pipelinesTue, 08 Sep 2026 12:00:00 GMTHow PII detector recall becomes residual identifiers, why Safe Harbor fields can leave unique records, how to measure recall, and what an audit trail holds.Expert and sensitive-domain data Guardrail classifiers under distribution shifthttps://reinforcedlabs.ai/research/guardrail-distribution-shifthttps://reinforcedlabs.ai/research/guardrail-distribution-shiftTue, 21 Jul 2026 12:00:00 GMTHow label, covariate and attack drift move guardrail precision and recall, when EM prior correction works, why small audits miss recall, and what to monitor.Guardrails and monitoring Is the chain of thought faithful? Tests, monitors and false negativeshttps://reinforcedlabs.ai/research/cot-faithfulnesshttps://reinforcedlabs.ai/research/cot-faithfulnessTue, 14 Jul 2026 12:00:00 GMTHow CoT faithfulness is tested, why reveal rates need a drift correction, how faithfulness sets a CoT monitor's miss rate, and what training against it does.Guardrails and monitoring Process vs outcome supervision: right answers, wrong reasoninghttps://reinforcedlabs.ai/research/process-vs-outcome-supervisionhttps://reinforcedlabs.ai/research/process-vs-outcome-supervisionTue, 07 Jul 2026 12:00:00 GMTWhy outcome reward models cannot reduce right-answer, wrong-reasoning rates, what step labels cost, and when PRM-guided search amplifies hidden errors.Training and reward Long-context evaluation beyond needle-in-a-haystack: tasks, position effects and effective context lengthhttps://reinforcedlabs.ai/research/long-context-evalshttps://reinforcedlabs.ai/research/long-context-evalsTue, 09 Jun 2026 12:00:00 GMTMulti-needle, aggregation and multi-hop tasks, position effects, precise definitions of effective context length, and how many trials it takes to measure it.Evaluation science Memorization and membership inference in fine-tuned models: attacks, canaries and what mitigations buyhttps://reinforcedlabs.ai/research/memorization-membership-inferencehttps://reinforcedlabs.ai/research/memorization-membership-inferenceTue, 28 Apr 2026 12:00:00 GMTLoss, reference, Min-K% and zlib attacks, canary exposure, and a real fine-tuning run showing how epochs and duplication drive leakage; DP-SGD trade-offs.Expert and sensitive-domain data Evaluating LLMs for financial work: tolerance, hallucinated figures and model risk validationhttps://reinforcedlabs.ai/research/finance-llm-evalshttps://reinforcedlabs.ai/research/finance-llm-evalsTue, 31 Mar 2026 12:00:00 GMTScoring numeric answers over filings: absolute vs relative tolerance, hallucinated figures, challenger models with McNemar, and CUSUM drift monitoring.Expert and sensitive-domain data Evaluating LLMs for clinical use: severity-weighted harm, escalation and small expert setshttps://reinforcedlabs.ai/research/medical-llm-evalshttps://reinforcedlabs.ai/research/medical-llm-evalsTue, 24 Feb 2026 12:00:00 GMTPhysician-written rubrics, severity-weighted harm, scored abstention and escalation, subgroup analysis and honest uncertainty for small clinical test sets.Expert and sensitive-domain data Rubric-based evaluation and rubric rewards: atomic criteria, weights and noisy gradershttps://reinforcedlabs.ai/research/rubric-evalshttps://reinforcedlabs.ai/research/rubric-evalsTue, 20 Jan 2026 12:00:00 GMTDesigning atomic rubric criteria, weighting them, measuring per-criterion grader reliability, and using rubrics as RL rewards without criterion gaming.Evaluation science What expertise buys in training and evaluation datahttps://reinforcedlabs.ai/research/expert-data-valuehttps://reinforcedlabs.ai/research/expert-data-valueTue, 09 Dec 2025 12:00:00 GMTLabel error by annotator tier, why majority vote misses systematic errors, expert adjudication, routing and the cost-quality frontier for reward-model data.Expert and sensitive-domain data Testing for sandbagging: elicitation, noise injection and the statistics of strategic underperformancehttps://reinforcedlabs.ai/research/sandbagging-detectionhttps://reinforcedlabs.ai/research/sandbagging-detectionTue, 02 Dec 2025 12:00:00 GMTTesting whether a model hides capability on evals: fine-tuning elicitation, password-locked models, noise injection, consistency checks and statistical tests.Evaluation science Over-refusal and under-refusal are one trade-off: measuring and calibrating bothhttps://reinforcedlabs.ai/research/over-refusal-calibrationhttps://reinforcedlabs.ai/research/over-refusal-calibrationTue, 28 Oct 2025 12:00:00 GMTHow to measure over-refusal and harmful compliance together: borderline prompt sets, sample sizes, label noise, contrastive synthetic data and the frontier.Guardrails and monitoring Attack success rate is a curve, not a number: power laws for Best-of-N jailbreakshttps://reinforcedlabs.ai/research/asr-power-lawshttps://reinforcedlabs.ai/research/asr-power-lawsTue, 16 Sep 2025 12:00:00 GMTWhy Best-of-N jailbreak success follows a power law, how to extrapolate ASR to large budgets, compare defences and bound rare successes. Seeded simulations.Red teaming and security Predicting benchmark accuracy from small models, with intervals that coverhttps://reinforcedlabs.ai/research/eval-scaling-predictionhttps://reinforcedlabs.ai/research/eval-scaling-predictionTue, 22 Jul 2025 12:00:00 GMTSigmoid fits on log-compute, emergence as a metric artifact, continuous metrics and prediction intervals for scaling forecasts, tested by seeded simulation.Evaluation science Reward hacking with verifiable rewards: measuring and hardening the verifierhttps://reinforcedlabs.ai/research/rlvr-reward-hackinghttps://reinforcedlabs.ai/research/rlvr-reward-hackingTue, 15 Jul 2025 12:00:00 GMTHow policies exploit RLVR verifiers, why a small false-positive rate lifts training reward while true accuracy falls, and how to audit and harden them.Training and reward Preference label noise: annotator agreement, Bradley-Terry reward models and how many labels per pairhttps://reinforcedlabs.ai/research/preference-label-noisehttps://reinforcedlabs.ai/research/preference-label-noiseTue, 08 Jul 2025 12:00:00 GMTWhat annotator agreement does and does not tell you, how flip noise affects Bradley-Terry reward models and DPO, and how many annotators per pair to pay for.Training and reward Evaluating tool-using agents: outcomes, trajectories, state diffs and pass^khttps://reinforcedlabs.ai/research/agent-trajectory-evalshttps://reinforcedlabs.ai/research/agent-trajectory-evalsTue, 06 May 2025 12:00:00 GMTOutcome versus trajectory scoring, partial credit, state-diff checks for collateral damage, environment flakiness and why pass^k collapses with task length.Evaluation science Calibration and abstention for LLMs: ECE pitfalls, Brier scores and risk-coverage curveshttps://reinforcedlabs.ai/research/calibration-selective-predictionhttps://reinforcedlabs.ai/research/calibration-selective-predictionTue, 25 Mar 2025 12:00:00 GMTWhy ECE misleads, what the Brier decomposition adds, token versus verbalised confidence, and how to set an abstention threshold with a risk you can defend.Evaluation science Evaluating long-reasoning models: pass@k, majority vote, seeds and token budgetshttps://reinforcedlabs.ai/research/reasoning-model-evalshttps://reinforcedlabs.ai/research/reasoning-model-evalsTue, 11 Feb 2025 12:00:00 GMTWhy pass@k and majority-vote accuracy diverge, how much seed noise a 30-problem benchmark carries, and how to report accuracy per token budget and cost.Evaluation science Reward model overoptimization: why gold reward peaks while the proxy keeps climbinghttps://reinforcedlabs.ai/research/reward-overoptimizationhttps://reinforcedlabs.ai/research/reward-overoptimizationTue, 17 Dec 2024 12:00:00 GMTWhy gold reward rises then falls as you optimise a reward model, how best-of-n and RL differ in KL, and what KL penalties and ensembles really fix.Training and reward Item response theory for LLM benchmarks: measuring more with fewer questionshttps://reinforcedlabs.ai/research/irt-for-llm-evalshttps://reinforcedlabs.ai/research/irt-for-llm-evalsMon, 09 Dec 2024 12:00:00 GMT1PL, 2PL and 3PL models for LLM benchmarks, Fisher information, informative and adaptive subsets, and a seeded simulation of how few items you need.Evaluation science Position and verbosity bias in LLM judges: swap-and-average and length-controlled win rateshttps://reinforcedlabs.ai/research/judge-bias-position-verbosityhttps://reinforcedlabs.ai/research/judge-bias-position-verbosityTue, 26 Nov 2024 12:00:00 GMTA pairwise judge model with position and length terms, why swapping order fixes one bias but not the other, and how length-controlled win rates work and fail.Evaluation science Benchmark contamination detection: what each detector misseshttps://reinforcedlabs.ai/research/contamination-detectionhttps://reinforcedlabs.ai/research/contamination-detectionTue, 22 Oct 2024 12:00:00 GMTn-gram overlap, Min-K% Prob, canaries, guided completion and order tests for benchmark contamination, plus a simulation of how paraphrased leaks evade them.Evaluation science Your benchmark's error bars are too narrow: clustered standard errors for LLM evalshttps://reinforcedlabs.ai/research/eval-clustered-standard-errorshttps://reinforcedlabs.ai/research/eval-clustered-standard-errorsTue, 17 Sep 2024 12:00:00 GMTQuestions that share passages, templates or prompts are not independent. Design effects, clustered standard errors and paired tests, simulated.Evaluation science