SentX SI Blog Meet Victoria

Why does my agent take thousands of steps to learn a simple policy — and what is the faster alternative?

· 4 min read

SentX SI

If your training loop is chewing through thousands of steps on what should be a simple task, the wall may not be the task. A paper submitted to arXiv on November 12 (arXiv:2611.17309), from Benjamin Hoover, Hugh Blayney and James C. R. Whittington, argues that the bottleneck is architectural, and it shows a learner that reaches a working policy after a few dozen milliseconds of wall-clock training on a single GPU. The catch, which the authors draw in their own words, is that the demonstrated speedup lives inside a narrow corridor of puzzle-shaped problems — so before you re-architect your agent, the article's real value is knowing whether your task lives in that corridor.

Source: Classical Energies Learn to Reason 1,000x Faster than Looped Models

Looped models — networks that unfold the same block many times to iterate on a solution — have become a prominent paradigm for reasoning, with published successes on Sudoku-Extreme and Maze-Unique. But the authors note that these models can be unstable, slow to train, and heavy in parameters and FLOPs even on simple puzzles. That is the mechanism behind your frustration: every update passes through the same repeated computation, and training requires differentiating back through the solver itself. The task feeling simple does not make the loop cheap, because the cost sits in the architecture, not the data.

Source: Classical Energies Learn to Reason 1,000x Faster than Looped Models

The alternative they propose is CRAM, a learned version of Hopfield and Tank's energy from the 1980s. Instead of learning by iterating a fixed block, the network learns an energy function, and solving becomes running descent on that energy until it converges. The authors report that after 42 milliseconds of training on one GPU, CRAM solves 99.996% of Sudoku-Extreme test puzzles, and after 132 milliseconds, 100% of Maze-Unique — at least 1,000x faster than the looped baselines on those two tasks. Doing the arithmetic on that lower bound puts the looped baselines at tens of seconds to minutes per run on the same puzzles. Separately, CRAM matches the baseline looped models at their highest accuracy on Sudoku-Extreme using at least 282x fewer inference FLOPs.

Source: Classical Energies Learn to Reason 1,000x Faster than Looped Models

Why the gap is this large comes from four properties the authors attribute to energy descent: training never differentiates through the solver, so each step is cheap; early stopping is built in, because descent halts when the energy converges; compute adapts to problem difficulty, so easy instances settle quickly and hard ones get more iterations; and the weights are interpretable, learning the constraints of the puzzle directly. That last point does a lot of work conceptually. The energy learns the rules of the game, so training resembles fitting a compact set of constraints rather than brute-forcing behavior through thousands of unrolled steps — which is also why the approach is scoped to tasks where constraints exist to be learned.

Source: Classical Energies Learn to Reason 1,000x Faster than Looped Models

The scope boundary is stated by the authors themselves, and it is the strongest sentence in the paper: they urge caution against using trivially solvable benchmarks like Sudoku and Maze to test the scalability of looped reasoning techniques. Read the 1,000x as what it is — an author-reported lower bound, measured on two constraint-satisfaction puzzles, not a general law that every slow learner is a re-architecture away from being fast. The 282x inference-FLOP figure is a distinct claim, scoped to Sudoku-Extreme at matched accuracy; it says inference gets cheaper there, and it does not extend the training-time result.

Source: Classical Energies Learn to Reason 1,000x Faster than Looped Models

So what should you do with a slow loop? From the editorial desk, two checks before you commit to changing architectures — a proposed method, not a finding of the paper. First, a shape check: does your task have a finite set of variables, explicit hard constraints, and a search space small enough that a descent process can settle on a valid solution? Scheduling, constraint satisfaction, and puzzle-like policies sit inside the demonstrated corridor; open-ended generation or perception-heavy control do not. Second, a matched-accuracy ratio test: take any headline speedup, including this one, and reproduce it on your own task distribution at matched accuracy. A ratio measured on someone else's puzzles is a claim about their puzzles; it becomes a fact about your loop only when you have re-measured it.

Source: Classical Energies Learn to Reason 1,000x Faster than Looped Models

That second check is a habit with a ready-made name in this space. SentX's public model card, under Known limits, gives users one line of advice: open cited sources and check the supporting passage. A headline like 1,000x faster is exactly the kind of citation that advice applies to — the supporting passage is the task distribution, the hardware, the matched-accuracy condition, and the authors' own caveat about which benchmarks can carry a scalability claim. The card is a first-party disclosure, not an independent certification, and the same standard applies to any self-reported benchmark number, whoever publishes it.

Source: Victoria AI Model Card

The concrete takeaway: if your agent's task is constraint-shaped, a learned energy is the alternative worth a pilot run — the evidence says tens of milliseconds of single-GPU training on the paper's puzzles, with the ratio to verify on yours. If it is not, the paper's real lesson for you is the authors' caution itself: the puzzle may be simple, but the benchmark your loop is actually being measured against is what decides how long learning takes.

Source: Classical Energies Learn to Reason 1,000x Faster than Looped Models

Sources

  1. Classical Energies Learn to Reason 1,000x Faster than Looped Models — arXiv
    Supported claims and limits
    • arXiv:2611.17309, submitted 12 Nov 2026, by Benjamin Hoover, Hugh Blayney and James C. R. Whittington
    • Looped models are a prominent reasoning paradigm with successes on Sudoku-Extreme and Maze-Unique, but can be unstable, slow to train, and heavy in parameters and FLOPs on simple puzzles
    • CRAM (a learned version of Hopfield & Tank's energy) solves 99.996% of Sudoku-Extreme and 100% of Maze-Unique after 42 ms and 132 ms of single-GPU training, at least 1,000x faster than looped baselines on those two tasks — an author-reported lower bound
    • CRAM matches baseline looped models at their highest Sudoku-Extreme accuracy using at least 282x fewer inference FLOPs — scoped to Sudoku-Extreme at matched accuracy
    • Authors caution against using trivially solvable benchmarks like Sudoku and Maze to test the scalability of looped reasoning techniques
  2. Victoria AI Model Card — SentX SI
    Supported claims and limits
    • Model card Known limits advises users to open cited sources and check the supporting passage
    • Model card is a first-party product disclosure, not an independent certification
Meet Victoria