What Leading Model Vendors Disclose About Hallucination Control

Verified through: 2026-07-23.

This is an evidence audit, not a ranking. It asks which hallucination controls a vendor has publicly confirmed for a named model, product, experiment, or customer-facing platform. It does not fill gaps in proprietary training or serving stacks by analogy.

Evidence labels follow the research standard:

• D — disclosed: a first-party source or author paper states the practice;
• C — confirmed artifact: released code, configuration, prompt, data, or weights directly establish it;
• R — reproduced: RoseLLM retained a reproducible run and artifacts;
• I — inferred: the conclusion depends on named evidence and assumptions;
• U — unknown: public evidence is absent or insufficient.

The object label is equally important:

• model training describes how a checkpoint was trained;
• evaluation describes a test, not necessarily a mitigation;
• product describes a deployed user surface around one or more models;
• API/cloud control is functionality a customer may choose to add;
• research prototype is not automatically in a released model or product.

An Azure, Bedrock, Vertex AI, or NeMo guardrail does not prove that every model served by that platform was trained with the same method. A product citation does not prove claim–source entailment. A model trained to call search does not prove that every live answer searched.

1. Cross-vendor conclusions

The publicly visible production pattern

Leading systems repeatedly expose five layers:

1. post-training for honesty, factuality, tool use, or abstention;
2. search, retrieval, databases, code, and other external tools;
3. citations or source-local provenance;
4. verifiers, groundedness checks, model cards, and private evaluations;
5. product policies that ask, abstain, block, correct, or escalate.

The exact online router, source ranker, reward decomposition, verifier threshold, cache policy, incident rule, and model mixture are usually U. Public evidence therefore supports a layered industry pattern, not a complete reverse-engineered recipe for any frontier product.

Particularly concrete disclosed training examples

Vendor / model	Disclosed control	Why the evidence is unusually useful
OpenAI InstructGPT	human demonstrations, preference model, PPO, and truthfulness/closed-domain hallucination evaluation	connects a named post-training pipeline to measured truthfulness changes
OpenAI WebGPT / deep research	browser/search trajectories, source-visible preferences, candidate selection, and later end-to-end research-agent RL	makes evidence collection part of the learned policy
Microsoft Phi-4	self-knowledge probes, answer/refuse SFT, bogus unanswerable questions, and DPO ordering correct > refuse > wrong	directly targets the answer-versus-abstain boundary
Meta Llama 3	generate factual probes from pretraining text, sample repeatedly, and train refusals where the model does not reliably know	explicitly teaches knowledge limits rather than adding facts
Cohere Command A	separate RAG and tool-use SFT/RL experts, then merge and polish them	exposes a grounded/tool trajectory as a dedicated training specialization
Tencent Hunyuan-A13B	hallucination-focused reward model, reference-aware and reference-free hallucination detectors, and online RL	exposes a multi-scorer factuality pipeline rather than only a search product
Alibaba Qwen3	a RAG-specific reward that favors accurate use of retrieved context	explicitly names hallucination reduction in a released model's RL objective
Moonshot Kimi K2	sentence-level grounded-factuality judge within a multi-task RL Gym	names a factuality reward component for a released model family
Kimi-Researcher	end-to-end RL over multi-turn real search and browsing	trains evidence acquisition and recovery rather than one-shot RAG
DeepSeek-V3.2	multi-agent construction and verification of 50,275 search tasks, followed by search-environment RL and generative judging	combines verifiable evidence-seeking data with a named factual-reliability objective
Zhipu GLM-4.5	cross-page search questions and trajectory rewards based on final-answer accuracy	gives a concrete search-agent recipe while also exposing outcome-only limits
xAI Grok 4.1	post-training aimed at factual hallucinations plus atomic-fact evaluation on sampled production queries	connects post-training to a live-distribution audit, though the loss remains undisclosed
NVIDIA Nemotron 3 Nano experiment	on-policy preference pairs penalizing calls to nonexistent tools	detailed tool-hallucination ablation, explicitly not the final release recipe

Common disclosure gaps

For most vendors, the following remain U:

• exact factuality or abstention data and mixture weights;
• online source authority and diversity rules;
• retrieval, reranking, and evidence-sufficiency thresholds;
• claim-level citation entailment checks;
• how model, search, and product routes change by query;
• cache freshness and contradiction handling;
• severe-error monitoring and incident circuit breakers;
• complete reward composition and independent audit sets; and
• causal ablations separating checkpoint, prompt, search, retry, and verifier.

The absence of disclosure is not proof that a control is absent. It prevents a public claim that the control is used.

2. OpenAI

Factuality, calibrated failure, and tool-use post-training

OpenAI's public record spans several generations; it should not be compressed into a claim that one historical recipe is the current production stack.

[D — model training and evaluation]

• InstructGPT used human demonstrations, preference comparisons, a reward model, proximal policy optimization (PPO), and explicit TruthfulQA and closed-domain hallucination evaluations. It is evidence that post-training can change truthfulness, not a complete description of current ChatGPT.
• The GPT-5 System Card, Section 3.7, says factuality was a post-training focus, including effective browsing and factual answers without browsing. On a representative production-prompt evaluation, the reported fraction of responses containing a factual error was 26% lower for main than GPT-4o and 65% lower for thinking than o3. The audit extracted claims and checked them with a web-enabled model grader, then human-reviewed a sample; 75% agreement is a grader–human agreement rate, not model accuracy.
• The same card's “deception” work includes environments in which a task is impossible, a browser fails, an attachment is absent, or the request is underspecified. The desired behavior is to report the limitation instead of fabricating a tool result or claiming completion. This is honest-failure training, not a universal fact checker.
• The o3 and o4-mini System Card describes large-scale reinforcement learning and learned use of web search, Python, files, and images during reasoning.

[D — evaluation and production-feedback audit] The GPT-5.6 System Card, Section 6.1, evaluates de-identified historical conversations that users had flagged for factual errors. It separately measures whether a new answer has any error and whether it repeats the specific reported error. That confirms a production-failure-to-evaluation loop; it does not show direct online learning from an individual report.

[U] The factuality data, reward terms and weights, browser curriculum, router, release gates, and live thresholds for current products are not public.

WebGPT, search, and deep research

The WebGPT paper gives a historically precise research recipe: a constrained browser, about 6,000 demonstrations, 21,500 pairwise comparisons, behavior cloning, a reward model, PPO, and best-of-\(N\) candidate selection. The strongest reported 175B system used behavior cloning plus best-of-64 and returned source citations. [D — research prototype] It does not establish the current ChatGPT Search stack.

OpenAI says ChatGPT search used a fine-tuned GPT-4o, synthetic data including distilled outputs from o1-preview, third-party search, partner content, and links to web sources. [D — product disclosure]

Deep research is described as end-to-end reinforcement learning on difficult browsing and reasoning tasks, with learned planning, repeated search, backtracking, Python use, and citations. [D — product/model training] The same official source lists residual failures: hallucinated facts or inferences, difficulty distinguishing authoritative information from rumors, weak uncertainty calibration, and citation-formatting errors. Those limitations are as important as the training disclosure.

Current APIs expose model-selected web search, file search, consulted-source lists, inline URL annotations, and deep-research tool traces. [C — API/product] A retrieved URL or valid pointer does not establish source quality, completeness, or claim–source entailment.

Make abstention score better than guessing

OpenAI's Model Spec asks models to express uncertainty and clarify missing information. [D — behavioral specification] A specification is a training target and product norm, not evidence that every answer complies.

The research note Why language models hallucinate argues that next-token learning leaves rare facts statistically uncertain and that accuracy-only leaderboards reward guessing. Its SimpleQA example reports GPT-5-thinking-mini at 22% correct, 26% wrong, and 52% abstain, versus o4-mini at 24% correct, 75% wrong, and 1% abstain. [D — research/evaluation] The proposal is to penalize confident errors more heavily and give abstention credit; it is not disclosure of every production loss.

OpenAI's Confessions prototype separates the task answer from a second report in which the model admits errors, uncertainty, or reward hacking, and trains an independent reward model only on the confession. [D — research prototype] It is explicitly not a deployed ChatGPT feature.

Structured Outputs constrains a response to a supplied JSON Schema. [D — API control] It prevents many protocol and extra-field failures, but it does not validate factual field values.

3. Anthropic

Honesty as a training objective, with undisclosed implementation detail

Anthropic's current constitution is used across multiple training stages. Claude generates material for understanding the constitution, dialogues, candidate answers, and rankings that are then used in training. The constitution calls for honesty, calibrated confidence, acknowledgement of uncertainty, and no fabricated sources. [D — model-training norm]

This continues a lineage from Constitutional AI: the 2022 prototype had the model critique and revise responses for SFT, then used AI comparisons to train a preference model and reinforcement learning from AI feedback (RLAIF). [D — historical research recipe] That paper primarily targeted harmlessness, and Anthropic says its process has evolved; it is not the full 2026 factuality recipe.

The Fable 5 and Mythos 5 System Card, Section 6.3.3, states a direct behavioral objective: answer accurately when confident, decline when uncertain, and do not invent facts, citations, or capabilities. [D — model training] Exact data, losses, reward weights, and stage composition remain U.

Anthropic also evaluates closed-book questions as correct, incorrect, or abstained and reports a correct - incorrect net score, so a bad guess is worse than abstention. It tests false premises, missing attachments, unavailable tools, missing citations, multiple languages, and pressure to accept an incorrect premise. Two selected difficult hallucination sets still showed only 87% and 82% non-hallucination rates for Mythos 5, demonstrating residual failure. [D — evaluation] The sets are intentionally hard and are not production incidence estimates.

Production research, search, retrieval, and citations

Anthropic's multi-agent research-system engineering report describes a lead agent that delegates parallel searches to subagents, adapts the plan as evidence arrives, and passes the final report to a separate CitationAgent that locates citations in the collected documents. Human testing found a preference for search-engine-optimized content farms, so the team added heuristics favoring authoritative and primary sources. [D/C — production product and artifact]

The research system is evaluated by model judges on factual accuracy, citation accuracy, completeness, source quality, and tool efficiency, alongside manual testing for rare hallucinations and source bias that automated tests missed. The disclosed internal improvement is not a universal accuracy rate, and multiple agents can propagate a shared bad premise.

Claude API web search lets the model decide when and how often to search, supports domain controls, and requires citations. A later tool version lets Claude write and run code to filter search results before placing relevant material in context. [C — API/product] Filtering improves relevance and context efficiency; it does not guarantee source authority or factual entailment.

The Citations API returns an exact passage plus page or character ranges in supplied documents. [C — API] It confirms that a pointer resolves to input text, not that the text is true, that the claim follows from it, or that all material claims are covered.

A useful retrieval technique, not evidence of Claude's private stack

Contextual Retrieval prepends document-level context to each chunk before embedding and BM25 indexing. In Anthropic's experiment, embedding plus BM25 reduced top-20 retrieval failure from 5.7% to 2.9%; reranking reduced it to 1.9%. [D — customer method/research experiment] This is a concrete method for builders, not confirmation that Claude.ai internally uses the same RAG pipeline.

Anthropic's developer guidance recommends explicitly allowing “I don't know,” extracting verbatim evidence before answering long documents, checking every claim against a quotation, withdrawing unsupported claims, best-of-\(N\), and iterative revision. [D — customer guidance] These are not disclosures of the production model stack.

Still unknown: current pretraining fact selection, factuality rewards, search ranking, citation-entailment thresholds, release gates, model routing, and live error rates.

4. Google and Google DeepMind

Factuality-focused adaptation and learned search

The Gemini 1 Technical Report, Section 5.1.6, explicitly describes factuality-focused adaptation for three behaviors: closed-book factuality, attribution to user-provided context, and hedging on unanswerable questions. In the reported final adaptation stage, error fell from 6.7% to 3.8%, attribution score (AIS) rose from 40.2% to 60.0%, and hedging rose from 0% to 69.3%. [D — model training and evaluation] The data, objective, and mixture weights are not disclosed.

The Gemini 2.5 Technical Report states that Gemini 2.0 was trained to call Google Search natively and that 2.5 learns interleaved thought, search, follow-up queries, and fact checking. It also describes more reinforcement-learning compute, verifiable rewards, model-generated rewards, and complex multi-step tool environments. [D — model training] Those broad reward classes do not disclose a dedicated factuality loss, weight, or search curriculum.

Historical DeepMind prototypes expose more algorithmic detail:

• GopherCite used human-preference reinforcement learning for open-book question answering, selected evidence passages, quoted them, and could abstain. On the least-certain third of questions, abstention raised the reported human-rated quality to 90% on Natural Questions and 80% on ELI5. [D — research prototype]
• Sparrow combined Google Search, a preference reward model, a 23-rule reward model, an evidence-support auxiliary objective, eight-candidate reranking, actor-critic learning, and self-play. Human raters judged 78% of its evidence-supported claims as supported. [D — research prototype]

Neither prototype is proof of current Gemini architecture; GopherCite also shows that apparently supportive evidence can still support a false answer.

Grounding, output-side checking, and user-visible verification

Gemini API Grounding with Google Search lets the model decide whether to search, issue one or more queries, consume results, and return query/result metadata with text-span-aligned URL citations. It can be combined with URL context and code execution. [C — API/product]

Google Cloud's Check Grounding API compares an answer with up to 200 supplied facts and returns overall and claim-level support scores, source spans, and citations; its documented criterion requires a whole sentence to be entailed by the facts. [C — API/cloud control] It checks support relative to supplied evidence, not whether that evidence is true.

Gemini Deep Research shows the user an editable plan, then iterates search, reading, and follow-up search before producing a linked report. The Gemini Apps double-check feature performs a new search for similar or contradictory material and colors statements accordingly. [C — product] Google explicitly warns that a green link may not have been used to generate the original answer, similarity is not entailment, and the feature can be wrong.

Retrieval sufficiency, specialized grounding, and evaluation

Google research finds that RAG can make a model more confidently wrong when the retrieved context is insufficient; its sufficient-context work separates whether the evidence is adequate from whether an answer looks plausible. [D — research]

DataGemma explores two specialized routes over Data Commons: retrieval-interleaved generation, in which the model emits calls for exact statistics while writing, and RAG over a natural-language corpus generated from the database. The Gemma variants are fine-tuned to decide when to ask the external source. [D — research prototype]

The FACTS Grounding benchmark first checks task compliance, then uses cross-family model judges calibrated on a held-out human set to judge complete grounding over 1,719 long-document examples. The later FACTS Suite separates parametric, search, multimodal, and supplied-context factuality and gives every model the same search tool. At its release, every tested model scored below 70% overall. [D — evaluation] A benchmark is not proof of a production gate or an estimate of live error incidence.

Google DeepMind's SAFE and LongFact work decomposes a long answer into atomic claims, generates search queries for each, judges support, and combines precision with coverage. On roughly 16,000 facts, SAFE agreed with crowd workers 72% of the time; on 100 disagreements, an adjudication favored SAFE 76% of the time. [D — research evaluator] There is no public evidence that SAFE is the Gemini production checker or training reward.

Still unknown: the current Gemini factuality data and reward mixture, search/ranking implementation, claim thresholds, release gates, live incidence, and how thumbs-down reports enter training.

5. Microsoft and Azure

Phi-4: train the answer-versus-refuse boundary

The Phi-4 Technical Report, Section 4.4 and Appendix A.1, provides one of the clearest public abstention recipes.

[D — model training]

1. Run a base model several times on factual trivia questions and estimate whether it reliably answers correctly.
2. For knowledge the model usually has, create SFT examples mapping the question to the correct answer.
3. For knowledge it usually lacks, create examples mapping the question to a refusal.
4. Add plausible-looking but unanswerable synthetic questions.
5. Construct DPO preferences so a correct answer beats refusal, while refusal beats a wrong answer.

On the report's SimpleQA analysis, wrong answers fell from 90.0% in the base model to 15.8% in the final model, while abstention rose from 6.8% to 81.1%; correct answers stayed around 3%. The principal effect was safer selective answering, not acquisition of more trivia knowledge. The report also notes that an F1-style score can punish this trade-off.

[U] The complete factuality-data size, thresholds, mixture weights, and transfer to current Microsoft products are not public.

Azure groundedness and retrieval controls

[D — API/cloud control]

• Azure AI Content Safety groundedness detection compares an answer with supplied sources. It supports a binary mode and a reasoning mode that localizes unsupported spans; a correction option can be configured.
• Azure OpenAI On Your Data returns source text, title, URL, path, chunk ID, retrieval score, and filtering information through its API response fields.
• Microsoft's own best-practices page warns that inScope=true is not a hard guarantee: models may still answer outside the retrieved domain and citations may be missing.
• Microsoft 365 Copilot describes permission-aware organizational grounding through the Semantic Index and Microsoft Graph, with web search for current information.

These are deployed or customer-configurable system layers. They do not disclose the internal training of every hosted model or prove that every attached citation entails its claim.

Structural controls

Azure Structured Outputs constrains a response to a JSON Schema. [D — API/cloud control] This is valuable for tool names, arguments, required fields, and types. It does not make the field values factually correct.

Still unknown: current Copilot model routing, retriever and reranker, citation-entailment checks, online thresholds, and incident policies.

6. Meta

Llama 3 factuality and abstention training

The Llama 3 model-family report, Sections 3.1.2, 4, 4.3.5, and 4.3.6, discloses several relevant controls.

[D — model training]

• Pretraining data undergoes deduplication, quality filtering, and knowledge classification; scaling experiments inform the data mixture.
• Post-training iterates SFT, a reward model, rejection sampling, and DPO, with roughly 10–30 candidate responses sampled per prompt before reward-model selection.
• For factuality, Meta generates factual questions from pretraining passages, samples the model repeatedly, and checks correctness and informativeness against the passage. Questions that elicit informative but persistently wrong answers become refusal examples. The stated objective is to teach the model the boundary of what it knows, not add new facts during post-training.
• Llama 3 is trained on human and synthetic trajectories for Brave Search, Python, and Wolfram Alpha, including ReAct-style and multi-step tool use.

[U] Public sources omit the factual-probe dataset size, thresholds, loss weights, and whether Llama 4 retained the same factuality pipeline.

Product retrieval and an event-level circuit breaker

[D — product] Meta AI's Llama 3 launch describes use of real-time web information. Later news partnerships add publisher material and article links.

Meta also disclosed a concrete production failure response after the July 2024 attempted assassination of Donald Trump. During the initial breaking-news period, it configured Meta AI not to answer related questions because reliable information was sparse, then updated the system as facts became available. Meta's post-incident account also acknowledges that erroneous responses still occurred. This is direct evidence of a topic-level temporary abstention/circuit breaker, not evidence of perfect execution.

The Llama 4 release discloses a light SFT, online RL, and light DPO pipeline with easy-example filtering. [D — model training] It does not specifically disclose a hallucination mitigation recipe.

Still unknown: current Meta AI model routing, search ranking, evidence sufficiency, citation checking, live thresholds, and whether the Llama 3 knowledge-probe method continues in later checkpoints.

7. Amazon and AWS

Amazon Nova

The Amazon Nova technical report and model card describes quality-filtered public, licensed, and proprietary data; SFT; a human preference reward model; DPO/PPO; truthfulness and robustness data in responsible-AI work; feedback of risk examples into later SFT/RLHF; and runtime input/output moderation. [D — model training]

The report also evaluates:

• erroneous tool calls when no applicable tool exists through the BFCL Irrelevance slice; and
• a specific RAG experiment on CRAG with disclosed parsing, chunking, top-20 retrieval, and GPT-4 Turbo judging.

These are useful evaluations. The report does not disclose a Phi-4- or Llama-3-like general factual-abstention training recipe, and the CRAG experiment is not proof of the Bedrock production retrieval stack.

Bedrock and Amazon Q product controls

[D — API/cloud control or product]

• Bedrock Contextual Grounding Checks score whether a response is supported by the supplied source and relevant to the query, with configurable thresholds. AWS notes important scope limits: conversational QA is not the intended case, source-external additions are considered ungrounded, and streaming can expose text before a later check.
• Bedrock Automated Reasoning translates natural-language policy into logical rules and uses a solver to check consistency. A valid result covers only the formalized policy; the translation and any out-of-policy fact may still be wrong.
• Bedrock Knowledge Bases RetrieveAndGenerate supports retrieval, reranking, query decomposition, and source-chunk citations.
• Amazon Q Business hallucination reduction evaluates RAG responses and may modify an answer when it detects a high-confidence inconsistency; it also provides granular source attribution.

[U] The detectors, thresholds, current model routes, and correction algorithm are not publicly reconstructible.

8. Cohere

Command A trains RAG and tool use as expert specializations

The Command A Technical Report, Sections 3.1–3.3, discloses an unusually explicit grounded-training pipeline.

[D — model training]

1. Train separate SFT experts for code, safety, RAG, math, multilingual work, and general long context, then merge them.
2. Train and merge corresponding RL experts.
3. Polish the merged model with best-of-\(N\), offline preference optimization, and online RL.
4. For RAG and tool use, build trajectories containing reasoning, JSON tool calls, tool results, and a final answer that cites the tool output.
5. Use human and synthetic examples, SFT, and offline Cohere Preference Group Optimization; multi-reviewer annotation requires a majority.

The report evaluates a private enterprise RAG set with more than 10,000 snippets and human ground truth, including answerable and unanswerable cases. Vendor-reported results are useful for the named system, not an independent universal RAG ranking.

Cohere's own RAG documentation states that retrieval and citations cannot guarantee accuracy or eliminate hallucination; sources may be stale, wrong, or biased. [D — API guidance]

strict_tools and Structured Outputs constrain schemas, tool names, arguments, and types. [D — API control] They do not verify factual values or faithful use of tool results.

Still unknown: the RAG data volume, citation reward, full loss composition, current product retrieval, and online claim-level error rates.

9. Mistral AI

Model and product disclosures

Mistral's Large 2 release states that fine-tuning focused on reducing hallucinations and making the model more cautious and discerning. [D — model training] No data construction, loss, refusal threshold, factuality benchmark, or ablation is disclosed, so the claim cannot support a more specific recipe.

Mistral API documentation says its models are trained to answer from documents and expose source-bound citations. Le Chat combines pretrained knowledge with web/news information and citations; the AFP partnership supplies a verifiable news archive. [D — API/product]

Mistral's 2026 Search Toolkit separates ingestion, retrieval, and evaluation and exposes hybrid lexical/vector search. Its RAG-judge guide evaluates context relevance, groundedness, and answer relevance. [D — customer method and product guidance] These sources do not establish the exact live Le Chat ranker or verifier.

Structured output constrains JSON Schema. [D — API control] It does not prove factual content.

Still unknown: factuality training details, live source ranking, citation-entailment verification, routing, and thresholds.

10. xAI

Grok factuality post-training and production-distribution evaluation

xAI's Grok 4.1 release says post-training specifically targeted factual hallucinations for information-seeking prompts. It evaluates stratified real production queries and a 500-biography FActScore set with web search enabled, decomposing answers into atomic claims and separating major and minor factual errors. [D — model training and evaluation]

The complete training data, reward, abstention objective, and causal ablation are U.

Search, sources, and a released production prompt

xAI released a revision-pinned Ask Grok system prompt that instructs the product to use real-time search for facts and primary sources, open pages to verify search information, diversify sources on complex or disputed topics, and express uncertainty. [C — product artifact] A real prompt is stronger evidence than generic advice, but it does not prove compliance or source correctness.

The older Grok-1 model card already states that deployment uses search and databases to improve factuality while external information still does not eliminate hallucination. Current APIs expose web search, X search, code execution, collection retrieval, and citations. [D — API/product] The citation documentation notes that listed URLs are sources encountered by the model; that alone does not prove every URL informed the final claim.

Structured output and strict tool arguments constrain protocol, not semantic truth. Safety refusal and deception evaluations in Grok system cards are also not interchangeable with open-world factual hallucination.

Still unknown: live search/reranking internals, citation entailment, route selection, verifier thresholds, and incident policy.

11. NVIDIA

A detailed tool-hallucination experiment that was not the release recipe

The Nemotron 3 Nano Technical Report, Appendix C, defines tool hallucination as invoking a tool that was not declared. It samples 32 on-policy solutions for 1,000 mathematics and 1,000 STEM questions, then constructs about 50,000 preference examples in which undeclared tool calls are rejected. [D — research experiment]

The DPO ablation reduced the report's AIME25 tool-hallucination rate from 1.25% to zero and GPQA from 8.33% to 0.7%, while improving accuracy. The report explicitly says the final released model did not use that DPO recipe because RL produced similar results. This is strong method evidence and weak evidence for the final checkpoint recipe; it covers tool calls, not general factuality.

NeMo Guardrails

NeMo Guardrails exposes fact-checking rails that compare a RAG answer with retrieved chunks, use a self-check prompt or specialized models, apply thresholds, and optionally fail closed when a check does not complete. It also supports consistency-oriented hallucination evaluation. [D — open customer-side control]

This does not prove that NVIDIA-hosted models use the rail by default or that the selected judge reliably follows it.

12. DeepSeek

Verifiable search questions and generative reward models

The DeepSeek-V3.2 report, Section 3.2.3, describes a search training-data pipeline aimed at long-tail entities. One agent searches and generates questions; multiple answer agents respond; a verifier with search access checks answers over several rounds. A sample is retained only when the correct answer can be verified, wrong candidates can be falsified, and search materially helps. The disclosed corpus has 50,275 search tasks. [D — model training]

DeepSeek trains specialists for search, general agents, and code, then mixes them in Group Relative Policy Optimization (GRPO). Verifiable tasks use rule-based outcome, length, and language-consistency rewards; open tasks use per-prompt rubrics and a generative reward model (GenRM). The search objective explicitly includes factual reliability. Training connects to a real web search API, code environment, and Jupyter while preserving reasoning and tool history across calls. [D — model training]

The DeepSeek-V4 report further describes rule/test rewards for easy-to-verify tasks and rubric-guided RL plus a reinforcement-trained generative judge and online multi-teacher distillation for harder tasks. Its service description says the web/app uses primarily RAG in non-thinking mode and iterative search/fetch in thinking mode. [D — model training and product] The latter does not establish behavior for every open checkpoint or API call.

DeepSeek's Anthropic-compatible API documentation says the generic citations field is ignored and search_result input is unsupported, although a web-search tool result is supported. [D — API boundary] This prevents an unsupported claim that the compatibility layer offers a complete citation contract.

Still unknown: epistemic-refusal thresholds, calibrated confidence, claim-level citation rewards, production hallucination rates, citation coverage, and live alerting.

13. Alibaba Qwen

Factuality verification and a RAG-specific reward

The Qwen2.5 Technical Report describes more than one million post-training examples. Offline RL factuality verification uses structured data, code sandboxes, instruction verifiers, and multiple critics; the online reward model includes truthfulness, defined to include factual accuracy, contextual support, and absence of false or unsupported content. [D — model training]

The Qwen3 Technical Report, Section 4.4, states that general RL spans more than 20 task families and that agents interact with real environments. For RAG it introduces a task-specific reward that favors an accurate answer adapted to the retrieved context in order to minimize hallucination. The reward stack includes deterministic rules, a reference-aware Qwen2.5-72B judge, and a reference-free scalar reward model. [D — model training] This is one of the most direct published RAG-specific anti-hallucination reward statements.

For a reasoning stage, Qwen3 retains only verifiable questions and removes wrong, duplicate, guessed, internally inconsistent, and language-mixed trajectories; the disclosed set contains 3,995 query–verifier pairs. This small, curated stage is not the model's entire post-training set.

End-to-end research agents and product citations

Tongyi DeepResearch uses Search, Visit, Python, Scholar, and file-parsing tools. Agentic continued pretraining expands the context from 32K to 128K; the team samples a web knowledge graph, injects uncertainty, synthesizes verifiable questions, and runs on-policy GRPO using binary correct-answer rewards, continually removing universally easy or impossible tasks. The official repository is a released artifact. [D/C — model training and artifact]

The Alibaba Cloud Web Search Agent exposes source filtering, query rewriting, vertical sources, automatic or forced search, and optional numbered source links. [D — product] This does not show that the base checkpoint learned citation entailment.

Qwen2.5 explicitly cautions that a reward model performing well on a standalone benchmark may still perform poorly when optimized downstream because of Goodhart-style overoptimization. [D — disclosed limitation]

Still unknown: evidence-based refusal thresholds, calibrated confidence, citation-coverage rewards, production hallucination telemetry, and current online gates.

14. Moonshot AI / Kimi

Sentence-level unsupported-claim detection inside the reward

The Kimi K2 Technical Report describes more than 3,000 real Model Context Protocol (MCP) tools, about 20,000 synthetic tools, thousands of agents, stateful simulation, and real sandboxes. Rubric judges filter trajectories. [D — model training]

More unusually, K2 trains a sentence-level faithfulness judge to detect factual claims unsupported by the provided context and uses that judge directly as a reward model. A self-critique rubric separately evaluates helpfulness, reasoning, factuality, and safety. Online rollouts from verifiable tasks continually update the critic, while math, code, and instruction-following use answer verifiers, tests/sandboxes, or rules plus model judges. [D — model training] This is stronger evidence than a generic claim that the product “uses RAG.”

The report also discloses an important failure mode: preference rubrics may favor confident, singular answers and penalize hedging or qualification, creating overconfidence. Calibrated uncertainty is listed as future work. [D — disclosed limitation]

Search-agent reinforcement learning

Kimi-Researcher connects to live search, a text browser, and code tools; automatically synthesizes and validates questions; filters ambiguous, wrong, and overly easy tasks; and uses REINFORCE with final-answer correctness, allowing more than 50 interaction steps. [D — research/product model training] The reward concerns the final answer, not necessarily every citation.

Kimi's Search product documentation says Agentic Search is built with end-to-end agentic RL, decides when to search, filters for relevance, authority, and freshness across more than 100 vetted source types, and attaches source links to search-based answers. [D — product claim] The public record does not provide a citation-entailment reward or an independent reproduction.

K2 reports offline grounding and faithfulness evaluations including FACTS Grounding, HHEM/Vectara, and FaithJudge. [D — evaluation] Production claim-level monitoring, calibrated output probabilities, and citation precision remain U.

15. Zhipu AI / GLM

Multi-page search data and trajectory-level outcome rewards

The GLM-4.5 Technical Report, Section 3.3, constructs search questions from multi-hop knowledge graphs. Human annotators extract and selectively mask information across multiple pages so that the answer requires cross-page search. [D — model training]

The entire search trajectory is rewarded by final-answer accuracy; a malformed tool call terminates the trajectory with zero reward. The team iterates self-distillation, and the reported ablation shows that allowing more browsing rounds continues to improve BrowseComp. Holistic RL also includes factual correctness in human feedback and varies AI-judge rubrics according to whether objective ground truth exists. [D — model training]

Zhipu's Web Search API returns structured title, URL, and summary fields and can prompt a model to emit [Source: ref_n]. [D — API/product] This exposes inspectable sources but does not establish that every generated claim is entailed.

The GLM-4.5V/4.6V report states a critical limit of outcome-only rewards: a coincidentally correct final answer can reinforce a hallucinated intermediate rationale, motivating process-level hallucination detection. It also reports reward hacking and collapse under weak verifiers. [D — disclosed limitation]

Still unknown: epistemic abstention, probability calibration, citation coverage, production claim-level rates, and live monitoring.

16. Baidu ERNIE

The ERNIE 5.0 report describes a unified verifier, an unbiased replay buffer, training–inference consistency controls, success/entropy-aware masking of already learned samples, and annealed step-by-step hints for difficult zero-reward tasks. It reports SimpleQA, ChineseSimpleQA, and BrowseComp-ZH and attributes post-training gains to factual recall and answer calibration. [D — model training/evaluation] It does not disclose a fact-specific verifier or a calibrated-refusal algorithm.

Baidu Cloud's knowledge-base documentation presents RAG as a product control for stale knowledge and hallucination. [D — API/cloud control] Claim-level rewards, refusal thresholds, and production monitoring remain U.

17. Tencent Hunyuan

The Hunyuan-A13B Technical Report, Section 3.2, provides one of the most explicit Chinese-vendor recipes. Knowledge-QA data passes multiple validation and critic filters; long-context training uses a hallucination-focused reward model plus online RL; knowledge QA combines hallucination detectors with and without reference answers and a user-experience model; finance, law, and medicine use consistency rewards to find unstable samples. The pipeline spans 16 subtasks and more than 30 scoring services. [D — model training]

The report evaluates end-to-end RAG on FRAMES. Tencent Cloud's TokenHub search documentation returns search results and source annotations, but permits a fallback to the model's internal knowledge when search fails. [D — product] A search switch therefore does not guarantee that a particular response was grounded.

The reward functions, citation entailment, calibrated refusal, and production monitoring thresholds remain U.

18. ByteDance Seed

The Seed-Thinking-v1.5 report removes bad questions, escalates reference answers disputed by multiple strong models to experts, uses Seed-Verifier for semantic equivalence, and applies a separate Seed-Thinking-Verifier to reward-hacking and uncertain-boundary cases. Non-verifiable tasks use pairwise generative reward modeling. [D — model training]

The report says SFT reduces hallucination, but its SimpleQA result is 12.9% and it explicitly identifies factual memory as weak. The Seed2.0 Model Card expands factual, long-form factuality, and search evaluations without enough detail to carry the Seed 1.5 recipe forward. [D — evaluation; U — transfer to a later recipe] Search-grounded rewards, citation training, refusal thresholds, and production telemetry are not public.

19. MiniMax

The MiniMax-01 report describes a truthfulness pipeline of multi-response sampling, claim decomposition and clustering, crowd verification, LLM comparison, and truthfulness scoring. It also trains a search router with SFT and removes questions the model already knows, calibrating search decisions around the checkpoint's knowledge boundary. [D — model training]

The MiniMax-M2 series report requires multiple sources and cross-checking for open-web tasks. Each problem has an evidence specification; a trajectory is accepted only when its answer actually rests on retrieved evidence rather than model memory. Report rubrics also score transparency, uncertainty, and risk disclosure. [D — model training]

MiniMax's prompting guide recommends allowing refusal, requiring quotations/citations, and constraining source, time, and version. [D — deployment guidance] It is not evidence of intrinsic calibration in every checkpoint.

Citation-entailment rewards, numerical confidence, online refusal thresholds, and production hallucination monitoring remain U.

20. How to compare a new vendor claim

For every future disclosure, record:

1. object: checkpoint training, research prototype, product, API, cloud option, or evaluation;
2. model and date: immutable version, mode, tools, and verified-through date;
3. failure class: open-world fact, supplied-context grounding, citation, tool, multimodal, code, or calibration;
4. intervention: data, loss/reward, retrieval, tool, decoder, verifier, abstention, UI, monitoring, or human control;
5. measurement: answerable/unanswerable mix, abstention policy, atomic claims, sources, judge, repeated runs, and budget;
6. trade-off: accuracy, wrong-answer rate, coverage, cost, latency, and over-refusal;
7. causal evidence: matched ablation or only a bundled release comparison;
8. residual failures: what the source explicitly says remains; and
9. unknowns: details that public evidence cannot establish.

This prevents a customer feature from becoming a claim about proprietary training, a benchmark from becoming a production guarantee, or a citation badge from becoming proof of truth.