The reigning recipe for "reasoning" in AI is brute scale: take a giant transformer, make it spit out long chains of thought, and hope the right answer falls out of all those tokens. Two 2025 papers quietly attack that assumption. Both show that a model small enough to fit on a phone — tens of millions of parameters, not billions — can outscore frontier LLMs on some of the hardest structured-reasoning benchmarks we have, by thinking in a loop rather than thinking out loud.

The first is the Hierarchical Reasoning Model (HRM) from Sapient Intelligence. The second is "Less is More: Recursive Reasoning with Tiny Networks" (TRM) from Alexia Jolicoeur-Martineau at Samsung SAIL Montréal, which strips HRM down even further and still wins. This is a deep dive into how both work and why "less is more" turns out to be literally true here.

HRM: two clocks, one forward pass

HRM is a ~27M-parameter model that does its reasoning in latent space rather than in text. Instead of emitting "thinking" tokens, it iteratively refines a hidden state over a handful of short internal "thinking" bursts and then reads the answer straight out of that state.

The architecture is explicitly brain-inspired, leaning on three ideas borrowed from cortical computation: hierarchical processing, temporal separation, and recurrent connectivity. Concretely, there are two coupled recurrent modules running at different timescales:

  • A high-level module (H) — the slow, abstract "planner".
  • A low-level module (L) — the fast "worker" doing rapid, detailed computation.

The two modules share and update a common hidden state rather than producing separate outputs. Around this sits an outer loop: produce a candidate prediction, score a "halt or continue" signal, and either submit the answer or run another refinement burst. It is essentially System-1 fast computation and System-2 deliberate planning braided together inside a single forward pass — no external scratchpad, no chain-of-thought supervision.

What HRM achieves

The headline is sample- and parameter-efficiency. Trained from scratch on roughly 1,000 examples, with no pre-training and no CoT data, HRM reports:

  • 40.3% on ARC-AGI-1 with only 27M parameters and a 900-token (30×30 grid) context.
  • 5% on ARC-AGI-2, edging out far larger models such as OpenAI o3-mini-high, DeepSeek R1 and Claude 3.7 on that benchmark.
  • Near-perfect accuracy on Sudoku-Extreme and on optimal pathfinding in 30×30 mazes, on tasks where standard CoT methods score essentially 0%.

The ARC Prize team independently re-tested HRM on their hidden semi-private set, which both grounded the result and prompted a closer look at why the architecture works — a question the next paper takes head-on.

TRM: "Less is More"

Jolicoeur-Martineau's TRM starts from a blunt observation: HRM is impressive but poorly understood, and its justification leans on biological analogy and a fixed-point (1-step implicit) gradient approximation that may not be doing what the story claims. So TRM throws the scaffolding out and asks how simple a recursive reasoner can be while still working — no two-timescale biological argument, no fixed-point theorems.

The answer is startlingly small: a single tiny network with just 2 layers and ~7M parameters. TRM keeps three things in play — the embedded input question x, a current answer y, and a latent "scratchpad" z — and recursively improves the answer:

# Tiny Recursive Model, conceptually
# x: embedded question, y: current answer, z: latent scratchpad
for step in range(N_sup):        # up to ~16 improvement steps
    for _ in range(n):           # recursive reasoning on the scratchpad
        z = net(x, y, z)         # refine latent given question + answer
    y = net(y, z)                # update the answer from the scratchpad
return y

That inner loop refines the latent z several times, then folds it back into a new answer y; the whole thing repeats. Each pass can catch and correct errors from the previous answer, so the model progressively polishes its guess instead of committing to one shot. Crucially, it does this with one small network reused everywhere, rather than HRM's two specialized modules.

Why tiny + recursive wins

The counter-intuitive part is that shrinking the network helped. With so little data, a bigger model overfits; a 2-layer net forced to recurse many times has to find a genuinely reusable update rule rather than memorising. Depth comes from iteration in time, not from stacking parameters — which is exactly what these puzzle benchmarks reward.

What TRM achieves

  • 45% on ARC-AGI-1 and 8% on ARC-AGI-2 with only ~7M parameters — beating LLMs like DeepSeek R1, o3-mini and Gemini 2.5 Pro while using under 0.01% of their parameters.
  • 87.4% on Sudoku-Extreme, versus HRM's 55.0%.
  • 85.3% on Maze-Hard, versus HRM's 74.5%.

And the cost is almost a footnote: the paper reports training for ARC-AGI at under $500.

HRM vs TRM, side by side

The trajectory is clean. HRM proved the surprising point — latent recursion lets a sub-30M-param model rival or beat giant LLMs on abstract puzzles. TRM then showed that most of HRM's machinery wasn't load-bearing:

  • Params: HRM ~27M (two modules) → TRM ~7M (one 2-layer net).
  • Justification: HRM's neuroscience + fixed-point gradient story → TRM's plain recursive refinement, no biology required.
  • Results: TRM is ahead on ARC-AGI-1 (45% vs 40.3%), ARC-AGI-2 (8% vs 5%), Sudoku-Extreme (87.4% vs 55.0%) and Maze-Hard (85.3% vs 74.5%).

Same core idea — iterate on a latent answer until it's right — minus the parts that turned out to be optional.

Implementation note: I'm building my own minimal version of these recursive reasoners — an interactive test you can run in the browser will follow here soon. The plan is a from- scratch TRM-style loop you can step through on a small Sudoku/Maze instance to watch the latent scratchpad converge.

The takeaway

Scaling parameters is not the only road to reasoning. On structured-puzzle benchmarks, a tiny network that loops — reading its own answer, refining a latent scratchpad, and trying again — can outpace models thousands of times its size. HRM made the case provocatively; TRM made it cheaply and simply. For anyone interested in efficient reasoning, "less is more" is no longer just a slogan here; it's a leaderboard.

Sources & Further Reading