We present Hito 2B, a 2-billion-parameter language model fine-tuned from Qwen3.5-2B using Hitonet's proprietary Progressive LoRA Merging (PLM) and Group Relative Policy Optimization (GRPO) pipelines. The model is trained to produce structured reasoning traces organized around a taxonomy of nested cognitive tags that scaffold comprehension, retrieval, verification, and commitment into explicit stages. We evaluate Hito 2B head-to-head against its base model under matched inference conditions and report large improvements on reasoning-heavy benchmarks (GSM8K +35 points, MATH-500 +10, ARC-Challenge +10, code synthesis +5), alongside qualitative capabilities atypical for the parameter count, including correct solutions to abstract ARC-AGI grid puzzles and AMC-level symbolic algebra. The training recipe also produces shorter reasoning traces that prevent unproductive expansion, and surfaces self-correction as a first-class observable behavior within a single response. This release comprises the model weights, the Cognitive Framework specification, curated example transcripts, and head-to-head benchmark results.
Research Findings
Three observations emerged during the development and evaluation of Hito 2B that, taken together, form the empirical basis of this release. We present them here in the order in which they influenced our interpretation of the data.
1. Cognitive tags function as structural constraints, not stylistic markers
Under a linear chain-of-thought formulation, reasoning is emitted as free-form natural language. The model retains the freedom to skip steps, pattern-match on surface features, and produce a plausible-looking derivation that does not implement the claimed derivation. At larger scales this failure mode is mitigated by raw capacity. At the 2B scale, it is prominent.
When reasoning is instead emitted as a sequence of tagged stages (<understand>, <recall>, <logic>, <verify>, <commit>), the model is required to allocate generation steps to each stage in turn. The <verify> tag in particular reserves output positions whose content must either confirm or contradict the preceding derivation; the <commit> tag reserves a terminal position that cannot wander. The tags are therefore not decorative. They constrain the policy distribution into a form that gates shortcut reasoning. Through Progressive LoRA Merging and reward shaping under GRPO, this constraint becomes an internalized behavior rather than a prompt-engineered overlay.
2. The framework reduces generation length and prevents unproductive expansion
During head-to-head evaluation against the base model, we consistently observed that Hito 2B arrives at committed answers with substantially shorter reasoning traces than the base Qwen3.5-2B under identical sampling conditions. Across the six benchmarks in our matched-conditions comparison, Hito's median thinking length was approximately one quarter of the base model's on equivalent prompts.
This efficiency appears to stem from the same structural constraint that shapes reasoning quality. Because the cognitive tags partition the reasoning into discrete stages with explicit transitions, the model cannot linger indefinitely in any one mode. The <commit> tag acts as a terminal gate that the policy has been trained to reach. The practical consequence is twofold: the model avoids the unproductive expansion into repetitive verification loops that we observed the base model entering on complex problems, and response latency improves meaningfully for equivalent task difficulty.
In tasks where the base model exhausted its generation budget without emitting a final answer, Hito typically converged on a committed answer well within budget. This is measurable in the evaluation data: the base model's median response time on hard items in our probe suite was approximately 33 seconds with a significant rate of empty responses, while Hito's median response time on the same prompts was below 10 seconds with consistent answer emission.
3. Self-correction is observable within a single generation
The combination of <doubt>, <verify>, and updated <commit> tags implements a first-class self-correction mechanism that operates within a single response rather than across turns. A <doubt> tag fires when the model identifies an inconsistency in its own reasoning. A <verify> tag then reserves generation positions to test the doubt. If the verification reveals an error, the <commit> tag updates with the revised answer before the response is emitted to the user.
This behavior is visible in the output rather than hidden, which has two consequences. First, downstream tooling can detect and surface self-correction events to users. Second, the training signal for producing this behavior is tractable, because the reward function can directly reinforce rollouts in which the sequence <doubt> → <verify> → revised <commit> leads to a verifier-correct final answer.
Interpretation
We do not claim that the Cognitive Framework introduces capabilities absent from the underlying weights. The more cautious claim is that the framework channels latent capability through a structured path that suppresses shortcut reasoning, bounds generation length, and exposes self-correction as an observable behavior. On tasks with significant decomposition structure (mathematical derivation, abstract rule inference, multi-step causal analysis), this yields measurable gains. On tasks that are primarily factual retrieval, the framework offers less leverage because there is no decomposition to enforce.
The scale of the observed gains on reasoning-heavy benchmarks, particularly the 35-point improvement on GSM8K under matched conditions, suggests that structured reasoning scaffolds are a meaningful intervention for small language models even when the underlying architecture and parameter count are held fixed.
Evaluation
All comparisons are between Hito 2B and Qwen3.5-2B (the base model we fine-tune from), using the identical harness, identical prompts, same 8k context, same 4000-token generation budget, and temperature 0. This is a matched-conditions comparison: the only variable is our training pipeline.
Hito 2B vs Qwen3.5-2B head-to-head
Benchmark
Category
Hito 2B
Qwen3.5-2B
Delta
GSM8K
math word problems
60%
25%
+35
MATH-500
competition math
15%
5%
+10
ARC-Challenge
scientific reasoning
75%
65%
+10
HumanEval-style
code synthesis
95%
90%
+5
Macro average (reasoning)
61.3%
46.3%
+15.0
Per-benchmark delta
Interpretation
Across four reasoning-heavy benchmarks covering mathematical derivation, abstract scientific reasoning, and code synthesis, Hito 2B shows consistent gains over its base of +5 to +35 points under matched sampling conditions. The magnitude of the gains correlates with the amount of decomposition each task requires: tasks demanding multi-step derivation (GSM8K, MATH-500) show the largest deltas, while tasks that are closer to pattern completion (code synthesis) show more modest but still consistent improvements.
Methodology transparency
The numbers above are from n=20 runs per benchmark under matched inference conditions. They are not a replication of each model's published benchmark, which uses each team's preferred harness, full test sets, and sometimes specialized evaluation modes. The matched-conditions comparison is the rigorous measurement of what our training recipe does to the same base model.
For context, Qwen3.5-2B's published numbers (from the official model card, full test sets) include MMLU-Pro 55.3 (non-thinking) / 66.5 (thinking), MMLU-Redux 69.2 / 79.6, IFEval 61.2 / 78.6, and GPQA 51.6 (thinking). Our head-to-head focuses on tasks where the effect of structured reasoning training is most visible.
Capabilities
The following are unedited capabilities demonstrations. Each links to the full transcript with commentary.
Abstract reasoning: ARC-AGI grid puzzles
Given two input/output example grids and a novel test input, Hito 2B infers the underlying transformation rule (e.g., 180-degree rotation) and applies it correctly. ARC-AGI is designed to measure fluid intelligence rather than memorization. Most small open models score near zero. Full transcript.
Symbolic mathematics: competition algebra
Given x + 1/x = 3, compute x³ + 1/x³. Hito derives the cube-expansion identity (x + 1/x)³ = x³ + 3(x + 1/x) + 1/x³ and commits to the numerical answer (18) through the Cognitive Framework's decomposition pattern. Full transcript.
Statistical reasoning: confounding variables
Given a study reporting that coffee drinkers live longer, Hito identifies the correlation-causation gap, enumerates alternative explanations (wealth, lifestyle, healthcare access), and articulates what an interventional study would need to establish. Full transcript.
Bayesian reasoning under base-rate neglect
On the classical disease-test problem (1% prevalence, 99% test accuracy, positive result), Hito reasons through the priors and arrives at approximately 50% actual probability of having the disease, rather than the naive 99%. This is the reasoning pattern most small models fail on. Full transcript.
Self-referential reasoning about its own nature
When a user comments "I can see your raw thinking," Hito engages with the metacognitive claim by drawing the distinction between simulated transparency (its emitted reasoning tokens) and actual access to internal states, without collapsing into either consciousness claims or dismissive boilerplate. Full transcript.
Deductive logic: case analysis
Hito solves a Knights-and-Knaves puzzle by enumerating each hypothesis (A is a knight; A is a knave), testing each for internal consistency against the observed statements, and committing to the consistent branch. Full transcript.
The Cognitive Framework
Reasoning in Hito 2B is emitted as structured, nested cognitive tags inside a <think>...</think> envelope. The committed answer follows after </think>.
The self-correction loop (<doubt> → <verify> → updated <commit>) operates within a single response, which makes the behavior observable in the output rather than hidden across turns.
Full specification, usage conventions, and integration patterns for downstream applications: COGNITIVE_FRAMEWORK.md.
Hito 2B is trained in two stages on top of Qwen3.5-2B:
Stage 1: Progressive LoRA Merging (PLM). Multiple rounds of LoRA fine-tuning on curated structured-reasoning data, with each round's adapter merged into the base before the next round begins. This produces a progressively specialized model that retains base capability while internalizing the Cognitive Framework grammar.
Stage 2: Group Relative Policy Optimization (GRPO) with Hitonet's custom reward formula and proprietary dataset. GRPO applies reward pressure that reinforces the behaviors demonstrated to produce capability gains in Stage 1. A reasoning-answer consistency signal explicitly rewards rollouts where the committed answer follows from the reasoning rather than being disconnected.
Both stages are Hitonet's proprietary methodology. The model weights are released under the Hitonet Community License. Replication of the training recipe for commercial use is separately licensed.
Licensing
Hito 2B is released under the Hitonet Community License (see LICENSE).
Use in revenue-generating products or services (direct or indirect)
Use inside an organization with more than $1M in annual revenue
Paid consulting, internal tooling, or client deliverables
Inclusion in training datasets sold or licensed to third parties
Embedding in hardware, SaaS, or paid API services
Commercial licensing
Hitonet offers commercial licenses scaled to use case and organization size, including bulk licenses, white-label deployments, and custom training partnerships. Pricing is reasonable for startups and scales with usage.
1@misc{hitonet2026hito2b,
2 title={Hito 2B: Structured Nested Reasoning in a 2-Billion-Parameter Model},
3 author={Hitonet},
4 year={2026},
5 publisher={HuggingFace},
6 howpublished={\url{https://huggingface.co/hitonet/hito-2b}}
7}
Acknowledgements
Hito 2B is built on top of Qwen3.5-2B by Alibaba Cloud. We thank the Qwen team for releasing a strong, capable base model that made this work possible. See NOTICE for full attribution.
