Native Reasoning: Internalized Inference

• Native Reasoning is a multidisciplinary paradigm where reasoning is inherently embedded within a system’s representational substrate and task-specific constraints.
• It spans formal semantics, communication protocols, and LLM training, emphasizing native language, cultural grounding, and symbolic computation.
• Its design enhances model interpretability, evaluation robustness on culturally authentic tasks, and efficient internal computation through native mechanisms.

Native reasoning is a cross-domain research term for reasoning that is anchored in the representational substrate, task ecology, or interaction loop native to a system, rather than appended as an external scaffold. Across recent work, the term is used for native human linguistic and cultural competence in evaluation benchmarks, model-internal symbolic or latent reasoning formats, retrieval and perception policies embedded inside inference, belief alignment in agent communication, and internal logics derived from formal calculi (Xu et al., 2021, Fabbri et al., 23 Jul 2025, Liu, 26 Aug 2025, Wang et al., 17 Sep 2025, Seo et al., 19 Feb 2026, Williams et al., 2021). The unifying contrast is not between reasoning and non-reasoning, but between reasoning that is intrinsic to the medium and reasoning that is translated, post-hoc, externally orchestrated, or verifier-dependent.

1. Conceptual scope and formal meanings

The literature does not assign a single formal definition to native reasoning. In formal semantics, "Native Type Theory" defines a native type system for a language $T$ as the internal higher-order dependent type theory of the presheaf topos $P(T) = [T^{op}, \mathbf{Set}]$, with type constructors derived from term constructors and with transition systems internalized so that structure and behavior can be reasoned about simultaneously (Williams et al., 2021). In this usage, nativeness is categorical and internal: one reasons inside the internal language canonically induced by the language itself.

In communication theory, "Semantics-Native Communication with Contextual Reasoning" models communication directly in an agent’s semantic space of actions, concepts, and symbols, and extends a fast System 1 semantic mapping with a System 2 contextual reasoning loop in which a speaker iteratively self-communicates with a virtual agent built on the physical listener’s unique way of coding semantics (Seo et al., 2021). In 6G research, "Reasoning-Native Agentic Communication for 6G" shifts the design objective from correct bits or correct semantics to aligned reasoning and actions across heterogeneous agents, introducing belief divergence, belief alignment, coordination drift, a Reasoning Coordination Plane, a shared ontology $K$, and bounded belief modeling through a Recursive Belief Engine (Seo et al., 19 Feb 2026).

In LLM training, "Native Reasoning Models: Training LLMs to Reason on Unverifiable Data" treats the reasoning trace $z$ as a latent variable and learns it from question-answer pairs alone, without human chains of thought and without external verifiers (Wang et al., 12 Feb 2026). In this setting, native reasoning means that the model discovers its own reasoning procedures because they increase the likelihood of the ground-truth answer. In architecture papers such as "Interpretable by AI Mother Tongue: Native Symbolic Reasoning in Neural Models," native reasoning means that the forward computation itself is mediated by discrete symbols learned via vector quantization, with symbol chains and gates constituting the model’s actual execution path rather than an auxiliary explanation layer (Liu, 26 Aug 2025).

Domain	Native object	Representative papers
Formal semantics	Internal language of the language’s own presheaf topos	(Williams et al., 2021)
Communication and networks	Beliefs, concepts, and reasoning alignment	(Seo et al., 2021, Seo et al., 19 Feb 2026)
LLM training and architecture	Latent traces, symbolic codebooks, native execution graphs	(Wang et al., 12 Feb 2026, Liu, 26 Aug 2025, Wu et al., 8 Dec 2025)
Evaluation	Original-language, culturally grounded, native-authored tasks	(Xu et al., 2021, Fabbri et al., 23 Jul 2025, Perez et al., 25 May 2025)

This dispersion of meanings suggests that native reasoning is best understood as a design principle: reasoning is "native" when it is expressed in the system’s own operative representational format and task-specific constraints, rather than reconstructed after the fact.

2. Native reasoning as human-native evaluation

One major line of work operationalizes native reasoning through benchmarks written by native speakers and grounded in native educational or cultural practice. "Native Chinese Reader" defines native-level Chinese reading comprehension through real Chinese high-school exam materials designed to assess the language proficiency of native Chinese youth (Xu et al., 2021). NCR contains 8390 documents with an average length of 1024 characters, 20477 questions, and four document types (00 modern-style without poetry, 11 classical-style without poetry, 22 classical poetry, 33 modern poetry). The best competition model achieved 59% test accuracy, while human evaluation showed an average accuracy of 79%, exposing a substantial human-model gap on long-context, classical, rhetorical, and commonsense-heavy Chinese reading.

MultiNRC: A Challenging and Native Multilingual Reasoning Evaluation Benchmark for LLMs" extends the native-evaluation idea to French, Spanish, and Chinese by using more than 1,000 native, linguistic and culturally grounded reasoning questions written by native speakers (Fabbri et al., 23 Jul 2025). The benchmark has 1,055 questions across four categories: language-specific linguistic reasoning, wordplay and riddles, cultural and tradition reasoning, and math reasoning with cultural relevance. None of the 14 evaluated models scores above 50% on MultiNRC, and most models perform substantially better in math reasoning in English compared to the original languages, with an average gain of +10%, indicating persistent challenges with culturally grounded knowledge. The benchmark explicitly contrasts original-language reasoning with translated evaluation and measures what it calls the "Mother Tongue Effect.

"AI4Math: A Native Spanish Benchmark for University-Level Mathematical Reasoning in LLMs" applies the same principle to university-level mathematics (Perez et al., 25 May 2025). It contains 105 original university-level math problems natively authored in Spanish across Algebra, Calculus, Geometry, Probability, Number Theory, Combinatorics, and Logic, each accompanied by a step-by-step human solution. The top models—o3 mini, DeepSeek R1 685B, and DeepSeek V3 685B—achieve over 70% accuracy, whereas LLaMA 3.3 70B and GPT-4o mini remain below 40%. Geometry, Combinatorics, and Probability remain persistently challenging for all models.

Across these benchmarks, native reasoning is not merely multilingual reasoning. It is reasoning that depends on original-language syntax, discourse conventions, genre, educational canon, idioms, wordplay, local calendars, local units, or classical registers. This suggests that translation-based evaluation can conceal failures that only appear when reasoning is embedded in the original linguistic and cultural substrate.

3. Native reasoning inside model computation

A second line of work makes reasoning native to the model’s internal computation rather than to the evaluation language. In "Interpretable by AI Mother Tongue," the model learns a codebook $C = \{c_k\}_{k=1}^K$, quantizes hidden states $x$ to the nearest codebook entry $z_q$, and uses the resulting discrete symbol both to route attention through a Symbolic Router and to inject symbolic content through an Intuition Gate (Liu, 26 Aug 2025). The symbol chain across layers is the model’s "thought chain," and the reasoning trace is therefore computation-relevant by design. The paper reports competitive accuracy on AG News alongside higher symbol purity, more calibrated gating, and verifiable reasoning traces.

"Native Reasoning Models: Training LLMs to Reason on Unverifiable Data" makes native reasoning a learning objective rather than an architectural bottleneck (Wang et al., 12 Feb 2026). The joint distribution is written as $$p_\theta(z, y \mid x) = p_\theta(z \mid x)p_\theta(y \mid x, z)$$, and training optimizes over latent reasoning traces $z$ using only question-answer pairs. The framework unifies prior verifier-free methods as different aggregation functions $f(c_1,\dots,c_T)$ over token probabilities and shows that reward choices emphasizing difficult tokens improve robustness to policy collapse. On Llama-3.1-8B, NRT-WS (-log p) reaches 56.2 average accuracy against 46.0 for SFT and 50.8 for RLPR.

"Native Parallel Reasoner" redefines nativeness at the execution-graph level (Wu et al., 8 Dec 2025). Reasoning is represented as a dependency graph $P(T) = [T^{op}, \mathbf{Set}]$0, and steps with disjoint parent sets are generated concurrently under explicit parallel attention masks and parallel positional encodings. The framework combines a self-distilled progressive training pipeline, Parallel-Aware Policy Optimization, and an NPR Engine built on SGLang/Multiverse. Across eight reasoning benchmarks, NPR trained on Qwen3-4B reports performance gains of up to 24.5%, inference speedups up to 4.6x, and 100% genuine parallel execution.

In domain adaptation, "CALM Before the STORM" argues that modern LRMs already possess native reflective reasoning patterns and that direct fine-tuning on traditional non-reflective datasets suppresses them (Tang et al., 5 Oct 2025). CALM uses an expert intervener to inject concise corrective hints into trajectories; these interventions modify fewer than 2.6% of generated tokens. The resulting model, STORM, is a 4B-parameter LRM that achieves a new state-of-the-art average accuracy of 68.9% across five optimization modeling benchmarks, matching the performance of a 671B LRM. In GUI automation, "InfiGUIAgent" similarly builds native reasoning through a two-stage supervised fine-tuning pipeline in which Stage 2 synthesizes hierarchical reasoning and expectation-reflection reasoning over GUI trajectories, enabling the agent to natively perform complex reasoning over screenshots and action histories (Liu et al., 8 Jan 2025).

Taken together, these papers define native reasoning as a property of the forward process, the training objective, or the execution graph itself. The common rejection is of post-hoc rationales that merely describe a black-box decision after the fact.

4. Native retrieval, active perception, and modality alignment

A third cluster of papers relocates reasoning into the mechanisms of retrieval and perception. "Improving Context Fidelity via Native Retrieval-Augmented Reasoning" introduces CARE, in which retrieval is internal to the model’s decoding process rather than delegated to an external retriever (Wang et al., 17 Sep 2025). The model generates > ... reasoning with <retrieval> ... </retrieval> spans that must be literal substrings of the provided context. Training combines 7,739 curated SFT instances from HotpotQA with GRPO reinforcement learning on DROP and MS MARCO using a reward $P(T) = [T^{op}, \mathbf{Set}]$1, with $P(T) = [T^{op}, \mathbf{Set}]$2, $P(T) = [T^{op}, \mathbf{Set}]$3, and $P(T) = [T^{op}, \mathbf{Set}]$4. On real-world QA, CARE raises LLaMA-3.1 8B average F1 from 44.54 to 59.83 and Qwen2.5 14B from 51.64 to 61.63; on CofCA, it improves LLaMA-3.1 8B from 48.14 to 61.83.

"Native Active Perception as Reasoning for Omni-Modal Understanding" makes perception itself a reasoning decision by formulating long-video understanding as a POMDP-based Observation–Thought–Action cycle (Xing et al., 17 Jun 2026). OmniAgent selectively queries frames, audio, or clips, distills them into persistent textual memory $P(T) = [T^{op}, \mathbf{Set}]$5, and reasons over that memory rather than over the full raw stream. It exhibits positive test-time scaling: on VideoMME-Long, accuracy rises from 53.4% at 6 turns to 59.6% at 52 turns. On LVBench, OmniAgent-7B reaches 50.5%, outperforming Qwen2.5-VL-72B at 47.3%, while using approximately 203 frames versus 768.

"Which Speech Representation Better Matches Text-Native Reasoning?" studies nativeness as alignment to a frozen text LLM’s reasoning dynamics (Ye et al., 10 Jun 2026). Holding the information rate fixed at 600 bits/s, the paper sweeps frame rates from 50 Hz down to 2.08 Hz and finds a consistent best regime for speech QA at 4.17 Hz with intermediate-layer representation alignment. This best regime is interpreted as the point where speech token length and semantic density most closely match the token-by-token reasoning dynamics internalized during text-only pretraining.

In remote sensing, "SkyNative" removes the pretrained visual encoder and feeds raw patch tokens directly into a unified autoregressive backbone with modality-aware decoupling (Yang et al., 18 May 2026). The paper introduces RSME-Bench to test whether answers are grounded in image evidence through progressive visual degradation and misleading textual prompts. On the blind test, SkyNative attains $P(T) = [T^{op}, \mathbf{Set}]$6 on RSME-Bench, whereas encoder-based baselines remain much lower, supporting the claim that native patch-level multimodal modeling increases visual reliance and robustness against prompt-induced language priors.

These works imply that a system may fail not because it lacks a reasoning head, but because retrieval, perception, or modality tokenization is not native to the reasoning dynamics of the backbone.

5. Native reasoning in communication and network control

In communication theory, native reasoning appears when the communication problem is defined over meanings, beliefs, and coordination rather than over signals alone. "Semantics-Native Communication with Contextual Reasoning" models actions, concepts, and symbols, and introduces a System 2 procedure in which agents iteratively align a mutual context by minimizing a loss built from cross-entropies or, in the $P(T) = [T^{op}, \mathbf{Set}]$7 case, KL divergences (Seo et al., 2021). The paper proves that the reliability of System 2 SNC increases with the number of meaningful concepts and that System 2 SNC significantly reduces the semantic representation length without compromising communication reliability.

"Reasoning-Native Agentic Communication for 6G" extends this idea to heterogeneous autonomous systems (Seo et al., 19 Feb 2026). Communication is triggered not by bit novelty or semantic novelty but by predicted misalignment in internal belief states. The framework adds a Reasoning Coordination Plane above the Data Delivery Plane, maintains approximate beliefs $P(T) = [T^{op}, \mathbf{Set}]$8, uses a shared ontology $P(T) = [T^{op}, \mathbf{Set}]$9, and transmits when $K$0. In collaborative humanoid manipulation, the reasoning-native setup yields Reasoning Alignment Score 94% versus 68% for semantic and 42% for classical communication; Decision Impact per Bit is 3.5× higher than semantic; Mutual Belief Stability over 1000 cycles stays above 91%, whereas semantic communication drops to approximately 60%. In autonomous intersection coordination, the agentic design achieves task success rates around 92% with communication overhead around 58% relative to the semantic baseline.

"6G-Bench" converts this architectural agenda into evaluation (Ferrag et al., 9 Feb 2026). It defines 30 decision-making tasks, derives 10,000 very-hard multiple-choice questions from 113,475 scenarios, and retains 3,722 as a high-confidence evaluation set after automated filtering and expert human validation. Across 22 foundation models, deterministic pass@1 ranges from 0.22 to 0.82; leading models reach intent and policy reasoning accuracy in the range 0.87–0.89; and pass@5 on selective reasoning-intensive tasks ranges from 0.20 to 0.91.

Here native reasoning means reasoning in network-native abstractions—intents, slices, MEC placement, trust, agent interoperability, distributed intelligence—rather than generic language reasoning performed after the network state has already been flattened into text.

6. Native world models, spatial structure, and open problems

In multimodal spatial reasoning, "N3D-VLM" argues that 2D VLMs lack intrinsic 3D object perception and therefore cannot reliably reason about front/behind, distance, or metric size (Wang et al., 18 Dec 2025). The model fuses image features with depth-derived 3D positional encodings, predicts 3D boxes of the form $K$1, and performs explicit 3D reasoning over those boxes. The data construction pipeline is reported as yielding over six times larger than the largest existing single-image 3D detection dataset. On N3D-Bench, N3D-VLM-7B attains 89.7% on open-ended questions and 92.1% on numerical questions; on RefCOCO-style grounding, it substantially improves projected IoU and 3D IoU over Qwen3-VL baselines. This use of explicit 3D state is native reasoning in the sense that geometric computation is performed over internal 3D structure rather than inferred heuristically from 2D appearance.

At the other end of abstraction, Native Type Theory provides a foundational account of nativeness: one reasons in the internal language induced by the language’s own syntax and semantics (Williams et al., 2021). This older formulation is conceptually important because it shows that the contemporary machine learning uses of "native reasoning" are not merely metaphorical. They often seek the same property: to move the reasoning process into the system’s own operative medium, whether that medium is a presheaf topos, a symbolic codebook, a retrieval trace, a 3D world model, or a belief-alignment plane.

The open problems are correspondingly diverse. NCR identifies long-context modeling, cross-register understanding, and pragmatic and rhetorical inference as unresolved obstacles to native-level Chinese reading (Xu et al., 2021). MultiNRC identifies persistent failures in native multilingual reasoning and substantial gains from English translation in culturally grounded math, especially for Spanish and Chinese, indicating unresolved language imbalance (Fabbri et al., 23 Jul 2025). NRT raises reward-design and policy-collapse issues for verifier-free reasoning (Wang et al., 12 Feb 2026). Reasoning-native 6G work highlights scalability of belief modeling, real-time constraints, ontology evolution, and adversarial reasoning (Seo et al., 19 Feb 2026). OmniAgent notes sequential OTA-loop latency and reliance on textual summaries (Xing et al., 17 Jun 2026). CARE reduces hallucinations but still allows residual retrieval irrelevance because any copied span from context can receive retrieval reward (Wang et al., 17 Sep 2025).

A plausible implication is that native reasoning will remain a heterogeneous research program rather than converging quickly to a single doctrine. Yet the surveyed work converges on one technical thesis: reasoning quality depends critically on whether the inference procedure, representational substrate, and evaluation setting are native to the underlying task structure. When they are not, performance may reflect translation shortcuts, post-hoc explanation, brittle prompt formats, or external retrieval artifacts rather than the target reasoning competence itself.