It is May 2026, and the field has stopped pretending hallucinations are going to disappear.

What has happened instead is more interesting. Researchers have spent the last eighteen months building an entire toolkit — fine-tuning methods, low-rank adaptation techniques, preference optimization frameworks, image-grounded decoders, multi-adapter compositions — designed not to eliminate hallucinations but to bound them. To calibrate models so that when they are uncertain, they say so. To constrain them so that when they answer, the answer is grounded in something verifiable.

This is a different mindset from "fix the model." It is closer to how engineers approach any probabilistic system: you cannot eliminate error. You measure it, you bound it, you make it visible. The question stops being is the model truthful and becomes is this model's error rate acceptable for this use case, given these guardrails, against this ground truth.

This article goes through what is actually working — for text, for images, across foundation models, language models, and specialized models. The math, the methods, the benchmarks. What companies have tried in the past, what they are doing now, and what the May 2026 state of the art actually looks like.

How We Got Here: A Short History

The first generation of attempts to reduce hallucinations was essentially "tell the model not to hallucinate." Prompt engineering. System messages. Chain-of-thought reasoning. Companies wrote elaborate instructions: "If you do not know the answer, say so." The model would say so — sometimes — and then continue to hallucinate confidently in the next sentence.

The second generation was Retrieval-Augmented Generation, introduced in production around 2023. Connect the model to a knowledge base. Retrieve relevant documents. Ground the response in retrieved context. This worked, and continues to work — but the 2025 Stanford HAI study showed even specialized legal AI tools built on RAG hallucinated more than 17% of the time. RAG reduces hallucination. It does not eliminate it. The retrieval can fail. The retrieved documents can be irrelevant. The model can ignore them.

The third generation, which is where we are now, accepts that hallucinations are structural and attacks them at multiple levels simultaneously: at training time through fine-tuning, at the parameter level through low-rank adaptation, at the preference level through DPO and its variants, at the decoding level through grounded inference, and at the architectural level through multi-adapter composition. These techniques are not alternatives. They compose.

Let me walk through the mathematics.

The Mathematics of the Problem

A language model parameterized by weights θ generates tokens by sampling from a probability distribution:

P(y | x; θ) = ∏ P(y_t | y_<t, x; θ)

Where x is the input prompt, y is the output sequence, and each token y_t is sampled conditioned on the input and the previously generated tokens. The model selects each next token by computing logits over the vocabulary and applying softmax to obtain probabilities, then sampling (or taking the argmax for greedy decoding).

The hallucination problem in this framework is precise: the model has been trained to maximize the likelihood of plausible-sounding text given its training distribution. When the input x is in the distribution it learned from, this works well. When x is out of distribution — or when the answer requires factual recall that the training did not provide — the model still produces high-probability tokens, but those tokens trace a path through the vocabulary that may have no relationship to truth.

The MIT 2025 finding sharpens this: models use more confident language when hallucinating than when stating facts. This is not a bug. It is a property of how probability flows. When the model has high entropy over plausible continuations, it tends to commit to whichever happens to win the sampling — and the language patterns associated with confident assertions ("definitely," "certainly," "without a doubt") are common in the training data of confident assertions, regardless of whether those assertions were correct.

To reduce hallucination, you need to do one of three things mathematically:

1. Change the distribution the model is sampling from (fine-tuning).
2. Add an auxiliary signal that down-weights non-factual continuations (preference optimization, grounded decoding).
3. Detect when the model's distribution is unreliable and abstain (calibration, refusal training).

The current state-of-the-art combines all three.

Method 1: LoRA — The Surgical Tool

Low-Rank Adaptation, introduced by Hu et al. in 2021, is now the workhorse of fine-tuning at scale.

The mathematics is elegant. Instead of fine-tuning all parameters of a weight matrix W ∈ ℝ^(d×k), LoRA freezes W and learns two small matrices A ∈ ℝ^(r×k) and B ∈ ℝ^(d×r), where r is much smaller than d or k:

W_new = W + ΔW = W + BA

The update ΔW is constrained to be rank r, dramatically reducing the number of trainable parameters. For LLaMa-3.1-70B, full fine-tuning requires approximately 1,120 GB of GPU memory for model states alone. LoRA with rank 16 introduces only 0.29% additional parameters, reducing GPU memory usage to 142 GB while preserving model quality.

Why this matters for hallucination reduction: LoRA lets you fine-tune cheaply on factuality-focused data without rebuilding the entire model. You can train one base model, then attach many small LoRA adapters — each calibrated to a different domain, each grounded in a different curated dataset.

┌─────────────────────────────────────────────────────────┐
│        Base Model (frozen, 70B params)                  │
│                                                         │
│  ┌──────────┐  ┌──────────┐  ┌──────────┐  ┌──────────┐ │
│  │ Medical  │  │  Legal   │  │ Coding   │  │ Customer │ │
│  │  LoRA    │  │  LoRA    │  │  LoRA    │  │   LoRA   │ │
│  │ (140M)   │  │ (140M)   │  │ (140M)   │  │  (140M)  │ │
│  └──────────┘  └──────────┘  └──────────┘  └──────────┘ │
└─────────────────────────────────────────────────────────┘
        ↑              ↑              ↑              ↑
    Grounded in    Grounded in    Grounded in    Grounded in
    medical docs   case law       codebase       help center

PREREQ-Tune, published at ICLR 2025, took this further. It uses a dual-LoRA architecture: one LoRA absorbs synthetic factual knowledge during a pre-training adaptation phase, and is then frozen. A second "skill" LoRA is trained on top to learn the actual task. The knowledge LoRA can be removed or swapped, leaving the skill LoRA generalizable. This disentangles what the model knows from what the model does, which is precisely the architectural separation hallucination research has been trying to achieve.

PREREQ-Tune significantly outperforms existing state-of-the-art hallucination reduction algorithms in improving LLM factuality across both short QA and long-form generation tasks. The framework enables a modular design with plug-and-play knowledge modules that control knowledge access and a skill module that works generically with any knowledge sources.

LoRA ensembles take a different angle. Train multiple LoRA adapters on the same task with different initializations or hyperparameters, then average their predictions. This produces measurably better-calibrated outputs — the ensemble's confidence more closely matches its actual accuracy. The few-shot baseline is well-calibrated but often wrong; a single fine-tuned LoRA is more accurate but overconfident in its wrong predictions; the LoRA ensemble provides improvements in both accuracy and calibration in terms of Expected Calibration Error.

Method 2: DPO and F-DPO — Preference at the Source

Direct Preference Optimization, introduced by Rafailov et al. in 2023, has largely replaced reinforcement learning from human feedback (RLHF) as the standard alignment method for production models in 2026.

The math of DPO is a clever reformulation. RLHF requires training a reward model on preference data, then using reinforcement learning (typically PPO) to optimize the policy against that reward. This is unstable, expensive, and sensitive to hyperparameters. DPO observes that you can derive an analytical relationship between the optimal policy and the reward function, and then optimize the policy directly against preference pairs without ever training a separate reward model.

The DPO loss for a preference pair (x, y_w, y_l) — where y_w is preferred and y_l is dispreferred — is:

L_DPO = -log σ(β · log[π_θ(y_w|x) / π_ref(y_w|x)] - β · log[π_θ(y_l|x) / π_ref(y_l|x)])

Where π_θ is the model being trained, π_ref is the reference model (typically the SFT-tuned base), σ is the sigmoid function, and β is a temperature parameter controlling how strongly the model diverges from the reference. The loss pushes the model to increase the relative likelihood of preferred responses while staying close to the reference model.

Why this matters for hallucination: standard DPO optimizes for whatever preferences humans express. If humans prefer fluent, confident-sounding responses over uncertain ones — which they do — DPO will train the model to be more fluent and more confident, whether or not its responses are factual. RLHF and DPO can therefore actively increase hallucination if the preference data rewards fluency over truth.

F-DPO, published in January 2026 and updated in April 2026, fixes this with a simple modification.

F-DPO uses binary factuality labels (factual vs. hallucinated). It applies a label-flipping transformation that corrects misordered preference pairs so the chosen response is never less factual than the rejected one. It adds a factuality-aware margin that emphasizes pairs with clear correctness differences, reducing to standard DPO when both responses share the same factuality.

The mathematical addition is a factuality margin term:

L_F-DPO = -log σ(β · [log ratio(y_w) - log ratio(y_l)] + α · m(y_w, y_l))

Where m(y_w, y_l) is the factuality margin — non-zero only when y_w and y_l differ in factuality — and α controls its strength.

The empirical results on Qwen3-8B are striking: F-DPO reduces hallucination rates by 5x — from 0.424 to 0.084 — while improving or preserving helpfulness across all seven evaluated models from 1B to 14B parameters. The method requires no auxiliary reward model, no token-level annotations, and no multi-stage training.

Hallucination Rate Reduction (Qwen3-8B)
═══════════════════════════════════════════════
Base model     ████████████████████████  0.424
Standard DPO   ███████████████████░░░░░  0.378
F-DPO          ████░░░░░░░░░░░░░░░░░░░░  0.084 (5x reduction)
                0         0.2        0.4

Method 3: Vision — When the Ground Truth Is the Image

Vision-Language Models face a specific version of the hallucination problem: object hallucination. The model describes an image but mentions objects that are not in it. This has been a persistent failure mode and is one of the most actively researched areas in 2025-2026.

The cleanest 2025 work is MARINE — Mitigating hallucinAtion via image-gRounded guIdaNcE — published as an ICML 2025 spotlight. The approach is training-free and API-free. MARINE incorporates a pre-trained object grounding vision encoder to extract object-level information from the image, then uses classifier-free guidance during text generation to bias the model toward grounded outputs.

Mathematically, classifier-free guidance modifies the logits during decoding:

logits_guided = logits_unconditional + γ · (logits_grounded - logits_unconditional)

Where γ is the guidance strength. The "grounded" logits are conditioned on the explicit object list extracted by the auxiliary vision encoder. The "unconditional" logits are the model's natural output. Higher γ pulls the model harder toward what is actually in the image.

MARINE works on multiple LVLM architectures, requires no fine-tuning, requires no API access to large models, and demonstrates significant reduction in object hallucination on POPE, MME, and CHAIR benchmarks. The auxiliary grounding model — typically DETIC or Grounding DINO — provides the ground truth that the LVLM is held against.

CHAIR-DPO takes a complementary approach for fine-tuning. The CHAIR (Caption Hallucination Assessment with Image Relevance) metric measures the fraction of mentioned objects that are not present in the image. CHAIR-DPO uses this metric to construct preference pairs: given two image captions, the one with the lower CHAIR_i score (fewer hallucinated objects) becomes the preferred response. The model is then fine-tuned with DPO on these preference pairs, becoming object-aware in the process.

The newer CoFi-Dec framework, published in January 2026 by researchers at the University of Minnesota and Lenovo, integrates multi-level visual processing: a coarse-to-fine attention pattern that mimics human visual processing, starting with scene-level understanding before focusing on details. This training-free decoding method significantly reduces both factual errors and semantic inconsistencies across challenging benchmarks.

A critical 2024 paper deserves mention: "Does Object Grounding Really Reduce Hallucination?" The authors offer the first systematic analysis of fine-grained object grounding on LVLM hallucination under an evaluation protocol that more realistically captures open-ended generation. Their finding: many earlier "reductions" relied on evaluation protocols using MSCOCO data extensively present in LVLM training. Under stricter evaluation, grounding objectives have little to no effect on object hallucination in open caption generation.

This is the kind of self-correction that mature fields do. The takeaway is not that grounding doesn't work — it does, but only when evaluated honestly, on data the model has not seen, in tasks that reflect actual deployment conditions.

Method 4: Multi-Adapter Composition

The most architecturally interesting development of the last year is the rise of multi-LoRA systems — frameworks that compose multiple specialized adapters at inference time.

LoraMap, published in 2024 and refined in 2025, creates dedicated reasoning LoRAs trained on fact-checking from different perspectives. Three LoRAs, each fine-tuned on a different reasoning dataset, then mapped to coordinate at inference. The paper shows LoraMap outperforms LoraHub (the previous standard for LoRA composition) with significantly fewer parameters than LoraConcat (which concatenates LoRAs and further fine-tunes them).

AutoRAG-LoRA, published in 2025, takes the integration further. It is a hallucination-aware RAG framework that combines:

• Automated prompt rewriting
• Hybrid retrieval (dense + sparse)
• LoRA-based generation adapters
• A dual-mode hallucination detection module (classifier-based plus self-reflective)
• A KL-regularized contrastive feedback correction loop that enables targeted fine-tuning on hallucination outputs

The KL regularization is the mathematically interesting part: it prevents the corrective fine-tuning from drifting the model too far from its original distribution, avoiding overfitting to edge hallucination cases. The model improves on factual alignment over time without degrading on the rest.

LoRAFusion, accepted at EuroSys 2026, focuses on the systems engineering of running multiple LoRAs efficiently — achieving up to 1.96× end-to-end speedup compared to Megatron-LM. This matters because the cost of running many small specialized adapters has historically been the bottleneck. When that cost drops, the practical viability of multi-adapter systems goes up.

The architectural picture that emerges:

                    ┌────────────────────────┐
                    │   User Query           │
                    └───────────┬────────────┘
                                │
                                ▼
                    ┌────────────────────────┐
                    │  Router / Classifier   │
                    │  (which domain?)        │
                    └───────────┬────────────┘
                                │
              ┌─────────────────┼─────────────────┐
              ▼                 ▼                 ▼
        ┌──────────┐      ┌──────────┐      ┌──────────┐
        │   RAG    │      │  LoRA-A  │      │  LoRA-B  │
        │ Retrieval│      │ (domain) │      │ (skill)  │
        └────┬─────┘      └─────┬────┘      └─────┬────┘
             │                  │                  │
             └──────────────────┼──────────────────┘
                                ▼
                    ┌────────────────────────┐
                    │   Base Model (frozen)   │
                    │   + Adapters Composed   │
                    └───────────┬────────────┘
                                │
                                ▼
                    ┌────────────────────────┐
                    │  Hallucination Detector │
                    │  (CLAP / MetaQA / etc.) │
                    └───────────┬────────────┘
                                │
                                ▼
                    ┌────────────────────────┐
                    │  Output (with refusal   │
                    │  if confidence too low) │
                    └────────────────────────┘

This is not the simple "one model, one fine-tune" architecture of 2023. This is a stack — specialized at each layer, calibrated at each transition.

What the Numbers Say

Here is the May 2026 picture, drawn from current benchmarks and published research:

Hallucination Rate by Method (TruthfulQA-style benchmarks)
═══════════════════════════════════════════════════════════
Base LLM (no intervention)        ████████████████████  60-80%
+ Prompt engineering              ███████████████░░░░░  45-65%
+ Standard RAG                    █████████░░░░░░░░░░░  17-35%
+ RAG + DPO alignment             ██████░░░░░░░░░░░░░░  10-20%
+ Full F-DPO + grounded RAG       ███░░░░░░░░░░░░░░░░░  5-12%
+ Multi-adapter + detection layer ██░░░░░░░░░░░░░░░░░░  3-8%
                                  0%      25%     50%

The trajectory is real. The numbers continue to drop. The asymptote — what the lowest achievable hallucination rate actually is — remains unknown, and may be non-zero by architectural necessity. But what was a 60-80% problem in 2023 is now, with proper engineering, a 3-8% problem in production deployments at the state of the art.

Three caveats are important.

First, these numbers are benchmark-dependent. A model that scores well on TruthfulQA can still fail on a specific niche domain it was never tuned for. Domain-specific evaluation matters more than general benchmarks.

Second, calibration matters as much as accuracy. A model that hallucinates 5% of the time and says "I'm not sure" appropriately is far more useful than a model that hallucinates 3% of the time and sounds equally confident in every response. The Expected Calibration Error metric is now widely reported alongside accuracy in serious evaluations.

Third, the cost-quality tradeoff is non-trivial. Full F-DPO plus grounded RAG plus multi-adapter inference is expensive. For many production use cases, the right answer is a careful single LoRA plus a good retrieval system — not the cutting edge of every available method stacked together.

What Companies Are Doing Now

The pattern across industries in 2026:

Healthcare has moved aggressively toward CHAIR-DPO-style grounded fine-tuning for clinical report generation, and toward MARINE-style image-grounded inference for radiology applications. RRG-DPO (Radiology Report Generation with DPO) specifically addresses the false-positive / false-negative tradeoff that classical RLHF struggled with in medical settings.

Legal tech companies have largely abandoned the "general LLM for legal work" approach and moved to retrieval-grounded systems with specialized adapters. The 17% hallucination rate Stanford reported in 2025 was the wakeup call. Modern legal AI products explicitly cite source paragraphs for every claim, abstain when retrieval confidence is below threshold, and run multi-model verification on high-stakes outputs.

Enterprise knowledge bases have converged on RAG + structured retrieval + small specialized adapters for each domain. The base model is rented from a provider (Claude, GPT, Gemini). The differentiation is in the adapter layer and the retrieval system.

Marketing and creative tools use grounded generation: a digital twin of the product (as in INDG/Grip), brand-approved claim libraries, controlled diffusion with depth and segmentation guidance. The AI is constrained to operate within structures defined by the brand team.

Coding assistants have moved to test-grounded generation: the LoRA is fine-tuned on code from the specific codebase, the output is validated against the existing test suite, and the model is calibrated to refuse rather than guess when uncertain.

The common pattern: in every domain, the production answer is not a single technique. It is a stack — a base model plus retrieval plus fine-tuning plus preference alignment plus inference-time guardrails plus structural constraints. Each component reduces hallucination at a different layer. Together they produce systems that are bounded, calibrated, and accountable in ways that single-model deployments never were.

The Honest Bottom Line

The mathematics of probabilistic generation imply that hallucination cannot be reduced to zero by training methods alone. The model is, fundamentally, a function that produces plausible continuations of its input. The plausibility space and the truth space are not the same space and cannot be made the same space by any amount of fine-tuning.

What the methods of 2025-2026 have shown is that the gap between plausibility and truth can be narrowed substantially — through fine-tuning that disentangles knowledge from skill (PREREQ-Tune), through preference learning that rewards factuality over fluency (F-DPO), through inference-time grounding (MARINE, CoFi-Dec), through multi-adapter composition (LoraMap, AutoRAG-LoRA), and through systems engineering that makes all of the above run at acceptable cost (LoRAFusion).

The shift in the industry is from chasing the dream of a truthful model to building the architecture of a truthful system. The model is a component. The system includes retrieval, validation, calibration, and structural constraints. The system has a hallucination budget. The system reports its uncertainty. The system abstains when it cannot answer with sufficient grounding.

This is how every other probabilistic technology has matured. Cars do not have zero accident rates. Networks do not have zero packet loss. Financial systems do not have zero fraud. Each of these is bounded by engineering — measurement, calibration, monitoring, controls — that converts a wild probabilistic process into a predictable production capability.

AI is now passing through the same transition. May 2026 is the moment where the engineering caught up with the ambition. The methods are real. The math is solid. The benchmarks are dropping. The systems are shipping.

The honest version of the next decade in AI is not "we solved hallucination." It is "we learned how to build systems that bound it well enough to deploy them responsibly in more and more domains." Which is, in the end, what engineering has always been.

By Soumia, a developer advocate focused on making complex infrastructure legible — through writing, speaking, and helping technical and non-technical audiences find common ground. I work at the intersection of cloud-native systems, AI, and editorial craft. — LinkedIn · Portfolio