# PAR Explained: Making Looped Language Models Track Longer Problems

Ant International and NUS researchers show why extra reasoning loops need a definition of progress, not just more computation.

By Paras Madan | 2026-10-21

Ant International and National University of Singapore researchers introduced Progress-Aligned Recurrence, or PAR, on October 20. It trains looped models for longer state-tracking problems, rather than launching a chatbot.

## What PAR changes

A looped model reuses a computation block, allowing more computation without adding model parameters.

Training only against a final answer leaves intermediate passes without an explicit assignment.

The [PAR paper](https://arxiv.org/abs/2610.27628) makes iteration k predict the state after k operations. It also anchors recurrent-state norms to their first-iteration values, controlling how the internal representation's scale changes as the block is reused.

## The 16-times length result

On the paper's swap-S5 task, a Qwen2.5-0.5B model trained on up to 32 moves achieves 95.5% mean full-state accuracy at 512 moves.

**Length here means the number of task operations.** It does not mean a 16-times larger context window, a general reasoning improvement of that size or a speedup.

That distinction is important for evaluating the claim. Solving a longer instance of a known rule system tests whether the learned computation remains stable. Answering an open-ended question can require discovering the relevant rules in the first place.

## Why stopping is part of the design

Variable computation raises another design question: when should the system answer?

For an engineer, the broader issue is whether an extra unit of compute has a predictable purpose. A system that sometimes corrects itself and sometimes overwrites a correct answer is difficult to budget or evaluate by loop count alone.

## What this research does not establish

The public release is a research paper. The paper and its [Hugging Face listing](https://huggingface.co/papers/2610.27628) do not provide a linked implementation or downloadable PAR checkpoint at the time of writing.

The experiments are also narrower than production reasoning workloads. They should not be read as evidence that this recipe fixes coding agents, factual question answering or arbitrary long-horizon planning.

A useful next test would ask whether the same progress structure can be defined for less orderly tasks. In a box-tracking problem, each operation advances a known state. In a debugging session, one investigation may reveal that the previous assumption was wrong. Progress is harder to label.

PAR makes a specific contribution to the computation question: extra loops become more meaningful when training defines what each loop should achieve. The next challenge is finding tasks where that definition remains useful outside a controlled sequence of moves.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders explain complex AI research and turn technical releases into clear stories that developers can understand and share.

[Book a 30-minute call](https://cal.com/paras-madan/30min).