Ouro-2.6B-Thinking: Efficient Iterative Reasoning

• Ouro-2.6B-Thinking is a 2.6B-parameter looped transformer that employs iterative latent computation by repeatedly reusing its layer stack, enabling efficient reasoning without explicit chain-of-thought generation.
• The model features adaptive depth allocation via a learned exit gate and entropy-regularized objectives, ensuring all loop steps are trained and used effectively to refine internal representations.
• Empirical results show that Ouro-2.6B-Thinking matches or exceeds larger non-looped models on challenging reasoning benchmarks, particularly excelling in tasks that require complex math and logical problem solving.

Searching arXiv for the cited papers to ground the article and verify identifiers. Ouro-2.6B-Thinking is a 2.6B-parameter reasoning-oriented member of the Ouro family of Looped LLMs (LoopLMs). It is not defined primarily by explicit chain-of-thought generation, but by iterative latent computation: a decoder-only transformer repeatedly reuses the same stack of layers over the same token sequence, thereby allocating additional internal computation before emitting tokens. In the published Ouro framework, the 2.6B base model is pretrained on 7.7T tokens with entropy-regularized adaptive depth and then further supervised on reasoning-heavy data to produce the “-Thinking” variant, which is reported to match or exceed substantially larger non-looped models on several hard reasoning benchmarks (Zhu et al., 29 Oct 2025).

1. Architectural identity

Ouro-2.6B-Thinking inherits the core Ouro-2.6B architecture: a decoder-only transformer with 48 transformer layers, hidden size $d_{\text{model}} = 2048$, multi-head attention, SwiGLU FFN, RoPE position embeddings, and vocabulary size 49,152 using the SmolLM2 tokenizer (Zhu et al., 29 Oct 2025). The defining departure from a standard transformer is that depth is realized by repeatedly applying the same stack of layers rather than traversing a single long stack of distinct layers.

Formally, if $H^L$ denotes the shared stack of $L$ layers, emb the token embedding, and lmhead the output head, the looped model with $t$ recurrent steps is written as

$$F^{(t)}(\cdot) = \mathrm{lmhead} \circ \underbrace{H^L \circ H^L \circ \cdots \circ H^L}_{t\ \text{times}} \circ \mathrm{emb}(\cdot).$$

Here, $F^{(1)}$ is an ordinary transformer, whereas $F^{(t)}$ for $t>1$ performs repeated latent refinement with shared parameters.

In the final Ouro configuration, the 2.6B model is trained with up to 4 loop steps. Because each loop applies the whole 48-layer stack once, Ouro-2.6B at 4 loops has effective depth comparable to a ~192-layer transformer, while retaining 2.6B parameters (Zhu et al., 29 Oct 2025). This parameter-sharing regime is central to the model’s claimed parameter efficiency.

A common misconception is to treat Ouro-2.6B-Thinking as merely a small model prompted to emit more reasoning text. That characterization is incomplete. The published account places reasoning inside the pretraining architecture itself, with explicit latent recurrence and learned halting, rather than solely in post-training prompting (Zhu et al., 29 Oct 2025).

2. Latent reasoning and adaptive depth

The loop mechanism defines a latent reasoning trajectory over hidden states. For an input sequence with embeddings $H^{(0)}$, the recurrent updates are

$$H^{(t)} = H^L(H^{(t-1)}) \qquad \text{for } t=1,\dots,T_{\max}.$$

Each $H^L$0 is a refined internal representation of the same sequence. The model can apply the LM head after each loop, so every loop produces a next-token distribution and a corresponding language-model loss $H^L$1 (Zhu et al., 29 Oct 2025).

Depth allocation is learned through an exit gate. At step $H^L$2,

$H^L$3

which induces a distribution over exit steps

$H^L$4

Pretraining uses an entropy-regularized objective

$H^L$5

with

$H^L$6

The uniform-prior interpretation is explicit in the Ouro paper: unlike geometric priors that bias toward shallow computation, the objective is designed so that all depths are trained and utilized more evenly (Zhu et al., 29 Oct 2025).

A second training stage sharpens halting behavior by supervising the gate with loss improvement between successive loop depths. For token $H^L$7 at step $H^L$8,

$H^L$9

and the target continuation probability is

$L$0

with $L$1 and $L$2. The adaptive gate loss then penalizes both underthinking and overthinking by comparing $L$3 against the predicted stop probability $L$4 (Zhu et al., 29 Oct 2025).

At inference time, Ouro uses Q-exit: it accumulates the exit-step CDF

$L$5

and stops at the first $L$6 for which $L$7. The threshold $L$8 becomes a direct accuracy-compute control (Zhu et al., 29 Oct 2025).

3. Pretraining pipeline and reasoning fine-tuning

The published Ouro training pipeline is unusually long for this parameter scale. The models are pretrained on 7.7T tokens across four major stages, with looping active throughout (Zhu et al., 29 Oct 2025).

During Stage 1a – Pre-train I (Exploration), training runs for 3T tokens with 8 recurrent steps, 4K sequence length, base learning rate $L$9, and batch size 4M $t$0 8M tokens. This regime exhibited instability. Stage 1b – Pre-train II (Stabilized) adds another 3T tokens, reduces recurrent steps to 4 for stability, and performs the “upcycling” that yields the 1.4B and 2.6B models. Stage 2 – CT Annealing contributes 1.4T tokens, raises context to 16K, emphasizes high-quality math and code data, and decays the learning rate to $t$1. Stage 3 – LongCT adds 20B tokens at 64K context. Stage 4 – Mid-training contributes 300B tokens total, combining 90B new high-quality QA and CoT data with replay from earlier stages, at 32K sequence length and learning rate $t$2 with cosine decay (Zhu et al., 29 Oct 2025).

The recurrent depth also required stabilization choices. The initial 8-loop exploration regime was later reduced to 4 loops, and the KL weight $t$3 for entropy regularization was reduced from 0.1 in Stage 1a to 0.05 in later stages to lessen gradient conflict between the task loss and the KL term (Zhu et al., 29 Oct 2025).

Ouro-2.6B-Thinking is the supervised reasoning variant built on top of that pretrained base. Its SFT stage uses 8.3M examples total, with Math: 3.5M, Code: 3.2M, Science: 808k, and Chat: 767k. Training runs for 2 epochs with maximum sequence length 32K, optimizer Adam, learning rate $t$4, and cosine decay (Zhu et al., 29 Oct 2025). The “-Thinking” suffix therefore denotes a reasoning-focused supervised stage layered on top of the pretrained LoopLM, not a separate architecture.

4. Capabilities and empirical profile

The published evaluations separate the base looped model from the reasoning-SFT model. For the base model, the reported pattern is parameter efficiency rather than raw scale: Ouro-2.6B with 4 loops is repeatedly compared against dense 3–12B baselines and is described as behaving like a strong 7–8B model on general reasoning, with especially strong mathematics (Zhu et al., 29 Oct 2025).

Benchmark	Ouro-2.6B R4	Selected comparison
MMLU	74.60	Qwen3-8B: 76.63
MMLU-Pro	55.73	Qwen3-8B: 53.72
BBH	80.46	Gemma3-12B: 78.41
GSM8K	81.58	Qwen3-8B: 83.09
MATH500	90.85	Gemma3-12B: 83.20
MBPP+	66.60	Qwen3-8B: 67.90

For the reasoning-SFT variant, Ouro-2.6B-Thinking is evaluated on difficult reasoning benchmarks rather than broad knowledge tasks.

Benchmark	Ouro-2.6B-Thinking R4	Selected comparison
AIME 2024	64.7	Qwen3-8B: 73.0
AIME 2025	50.3	Qwen3-8B: 66.7
OlympiadBench	76.4	Qwen3-8B: 75.3
BeyondAIME	39.0	Qwen3-8B: 38.0
SuperGPQA	53.7	Qwen3-4B: 51.9

These results support a specific interpretation: the model’s advantage is concentrated in reasoning-heavy domains rather than in general factual storage (Zhu et al., 29 Oct 2025). That interpretation is reinforced by synthetic analyses in the same work. In the Capo synthetic biography memorization task, both looped and non-looped models plateau at about 2 bits/parameter, indicating that looping does not increase knowledge capacity. In Mano, a modular arithmetic tree-evaluation task, looped models with comparable parameter count and often comparable FLOPs dramatically outperform non-looped baselines; at difficulty $t$5, a base 12×1 model achieves 34.8%, while a looped 3×4 model reaches 92.2%. In a synthetic 3-hop QA task, models with more loops require fewer unique training examples to learn the reasoning pattern (Zhu et al., 29 Oct 2025).

The paper’s own summary is therefore precise: Ouro’s advantage stems not from increased knowledge capacity, but from superior knowledge manipulation capabilities (Zhu et al., 29 Oct 2025).

5. Reinforcement learning and RLTT

A later study treats Ouro-2.6B-Thinking as the canonical testbed for reinforcement learning in LoopLMs and argues that terminal-only RL objectives are structurally mismatched to looped computation (Jonathan et al., 11 Feb 2026). In that formulation, for a response $t$6, each loop $t$7 produces a per-token latent thought distribution

$t$8

but sampling uses only the final loop: $t$9 The paper fixes $$F^{(t)}(\cdot) = \mathrm{lmhead} \circ \underbrace{H^L \circ H^L \circ \cdots \circ H^L}_{t\ \text{times}} \circ \mathrm{emb}(\cdot).$$0 during RL.

The baseline is GRPO applied only to the terminal loop. The proposed alternative, RLTT (Reward Latent Thought Trajectories), replaces terminal-only credit assignment with a weighted sum over loop-level log-probabilities: $$F^{(t)}(\cdot) = \mathrm{lmhead} \circ \underbrace{H^L \circ H^L \circ \cdots \circ H^L}_{t\ \text{times}} \circ \mathrm{emb}(\cdot).$$1 subject to $$F^{(t)}(\cdot) = \mathrm{lmhead} \circ \underbrace{H^L \circ H^L \circ \cdots \circ H^L}_{t\ \text{times}} \circ \mathrm{emb}(\cdot).$$2 and $$F^{(t)}(\cdot) = \mathrm{lmhead} \circ \underbrace{H^L \circ H^L \circ \cdots \circ H^L}_{t\ \text{times}} \circ \mathrm{emb}(\cdot).$$3. A KL penalty is applied to the frozen reference policy at the terminal loop. The main experiments use exit-probability weights $$F^{(t)}(\cdot) = \mathrm{lmhead} \circ \underbrace{H^L \circ H^L \circ \cdots \circ H^L}_{t\ \text{times}} \circ \mathrm{emb}(\cdot).$$4 (Jonathan et al., 11 Feb 2026).

Training is math-only RLVR on the MATH training split, with binary $$F^{(t)}(\cdot) = \mathrm{lmhead} \circ \underbrace{H^L \circ H^L \circ \cdots \circ H^L}_{t\ \text{times}} \circ \mathrm{emb}(\cdot).$$5 reward, 4× H200 140GB, 140 RL optimization steps, 32 prompts per step, $$F^{(t)}(\cdot) = \mathrm{lmhead} \circ \underbrace{H^L \circ H^L \circ \cdots \circ H^L}_{t\ \text{times}} \circ \mathrm{emb}(\cdot).$$6 rollouts per prompt, max generation length 2048, AdamW 8-bit, $$F^{(t)}(\cdot) = \mathrm{lmhead} \circ \underbrace{H^L \circ H^L \circ \cdots \circ H^L}_{t\ \text{times}} \circ \mathrm{emb}(\cdot).$$7, $$F^{(t)}(\cdot) = \mathrm{lmhead} \circ \underbrace{H^L \circ H^L \circ \cdots \circ H^L}_{t\ \text{times}} \circ \mathrm{emb}(\cdot).$$8, learning rate $$F^{(t)}(\cdot) = \mathrm{lmhead} \circ \underbrace{H^L \circ H^L \circ \cdots \circ H^L}_{t\ \text{times}} \circ \mathrm{emb}(\cdot).$$9, weight decay 0.1, max gradient norm 0.1, gradient accumulation 2, BF16, and KL coefficient $F^{(1)}$0 (Jonathan et al., 11 Feb 2026).

The reported gains over GRPO are large under matched training and inference conditions.

Ouro variant	MATH-500	AIME24	BeyondAIME	GSM8K	Math Avg.
Ouro-2.6B-Thinking	67.8	13.3	5.0	58.5	36.2
+ SFT	58.2	13.3	6.0	59.6	34.3
+ GRPO	71.6	16.7	6.0	59.7	38.5
+ RLTT	86.0	33.3	16.0	94.0	57.3

The paper emphasizes the deltas against GRPO: +14.4 on MATH-500, +16.6 on AIME24, +10.0 on BeyondAIME, +34.3 on GSM8K, and +18.8 on math average (Jonathan et al., 11 Feb 2026). Transfer outside math is also reported despite math-only RL training: on ARC-C, MMLU-ST, GPQA, and MBPP, the non-math average rises from 65.2 under GRPO to 71.8 under RLTT, with GPQA improving by +18.7 points (Jonathan et al., 11 Feb 2026).

A further claim is qualitative rather than purely score-based: RLTT produces shorter responses, higher early reward, lower terminal-loop entropy, improved per-loop accuracy at 1–4 loops, and higher GSNR on the hardest math tasks. Statistical tests under stochastic decoding report $F^{(1)}$1 for RLTT $F^{(1)}$2 GRPO on all math and non-math benchmarks (Jonathan et al., 11 Feb 2026).

This post-training result is important because it resolves an apparent tension in the earlier Ouro report, which had noted that RLVR on top of LoopLM showed no improvement over SFT under its infrastructure constraints (Zhu et al., 29 Oct 2025). RLTT suggests that the bottleneck was not merely “more RL”, but the mismatch between terminal-only RL objectives and multi-loop latent computation (Jonathan et al., 11 Feb 2026).

6. Internal states and relational preference encoding

Ouro-2.6B-Thinking has also been studied as an object of representation analysis rather than only as a generator. In that setting, it is described as a 2.6B-parameter looped transformer with iterative refinement, maximum 4 loop steps, hidden dimension 2048, and a dynamic early-exit mechanism whose default threshold 1.0 effectively disables early exit, while 0.87 was used for feature extraction (Kirin, 10 Apr 2026).

At loop step $F^{(1)}$3, the hidden state tensor is

$F^{(1)}$4

The study extracts the list $F^{(1)}$5 and uses lightweight evaluator heads of about 5M parameters to predict human preference on Anthropic HH-RLHF while keeping the base model completely frozen (Kirin, 10 Apr 2026).

The central empirical result is that preference is encoded predominantly relationally rather than independently. The best pairwise evaluator reaches 95.2% test accuracy on 8,552 unseen examples. A linear probe on pairwise differences reaches 84.5%, while the best nonlinear independent evaluator reaches only 65% test accuracy. A linear independent classifier scores 21.75%, below chance and with inverted polarity (Kirin, 10 Apr 2026).

Those numbers are backed by architectural comparisons. The best pairwise evaluator computes loop-wise pooled representations for chosen and rejected responses, forms difference vectors $F^{(1)}$6, applies bias-free LayerNorm to preserve antisymmetry, projects to 512 dimensions, processes the four-step sequence with a 2-layer GRU of hidden size 512, and then scores the pair with a small MLP (Kirin, 10 Apr 2026). The conclusion is not merely that preference is present, but that it is far more accessible from the relation between two trajectories than from an absolute scalar score on one trajectory.

The paper interprets the evaluator as a model-internal consistency probe rather than a direct proxy for human consensus. It also documents a methodological caution: a cosine learning-rate dead zone at epoch 2 accidentally preserved the generalization peak, and test accuracy then degraded from 95.2% at epoch 2 to 62.4% by epoch 5. The authors accordingly propose the flip test—explicitly evaluating both $F^{(1)}$7 and $F^{(1)}$8—as a mandatory diagnostic for pairwise preference evaluators (Kirin, 10 Apr 2026).

For understanding Ouro-2.6B-Thinking itself, this analysis implies that loop trajectories are not only a mechanism for next-token refinement. They also expose structured value-relevant geometry that can be read by small external heads without modifying the base model (Kirin, 10 Apr 2026).

7. Instruction-following trade-offs, conceptual antecedents, and limitations

A broader question is how a built-in thinking mode interacts with instruction following. A comparative study on Qwen3 and Hunyuan models does not evaluate Ouro directly, but its findings are relevant to any compact reasoning model with a native thinking mode. Aggregate pass-rate changes on IFEval are small, yet 10–20% of prompts switch between pass and fail across Thinking ON/OFF modes, indicating that thinking redistributes errors rather than uniformly helping or hurting (Kumar, 8 Jun 2026).

The most stable pattern is a split between Planning and Precision constraints. Under the Qwen3-derived grouping, the pooled class deltas are

$F^{(1)}$9

After matched-length control, the Precision drop shrinks but remains negative, while the Planning gain persists (Kumar, 8 Jun 2026). This suggests that built-in reasoning tends to help global counting, structure, and coordination, while harming exact local form such as casing, punctuation bans, or exact replication. For Ouro-2.6B-Thinking, this is a plausible implication rather than a direct measurement, but it is consistent with the broader trade-off between deeper latent computation and strict token-level control.

Conceptually, Ouro also belongs to a longer line of work on extra computation steps in language modeling. An early proof-of-concept introduced thinking tokens—special <T> symbols inserted after each word in an LSTM LLM—to give the recurrent state additional updates before the next normal token (Herel et al., 2024). That mechanism produced local improvements on some complex arithmetic or numerically rich sentences, although global perplexity often did not improve and worsened when too many thinking tokens were inserted. The relationship is conceptual rather than architectural: Ouro realizes extra internal computation through looped transformer depth and adaptive halting, not through visible placeholder tokens (Herel et al., 2024).

Several limitations recur across the Ouro literature. Performance degrades when extrapolating beyond the trained loop depth $F^{(t)}$0, even though safety scores may continue to improve with more loops; the initial 8-loop pretraining regime proved unstable; RLTT increases memory pressure because per-loop log-probabilities must be stored; and the main RLTT study keeps fixed $F^{(t)}$1 during both training and inference despite Ouro’s dynamic early exit capability (Zhu et al., 29 Oct 2025, Jonathan et al., 11 Feb 2026). Another limitation is scope: the strongest RLTT evidence is math-only during training, and the strongest instruction-following evidence comes from non-Ouro model families (Jonathan et al., 11 Feb 2026, Kumar, 8 Jun 2026).

Within those constraints, Ouro-2.6B-Thinking is best understood as a compact looped transformer in which reasoning is implemented as iterative latent refinement, supervised during pretraining and sharpened by later post-training. Its significance lies less in explicit verbal chain-of-thought than in the claim that a 2.6B model can trade parameter count for recurrent latent depth, yielding 8B-class reasoning behavior, structured internal preference representations, and reinforcement-learning gains when the full latent trajectory is brought into the objective (Zhu et al., 29 Oct 2025, Jonathan et al., 11 Feb 2026, Kirin, 10 Apr 2026).