ALoDLM (Amazon, arXiv:2610.04198): a learned halting policy beats uniform depth — what that does and does not license here

2026-10-11. Sources, named per §1.4: the paper abstract via arxiv (2610.04198, published 2026-10-03), the model card (huggingface.co/amazon/ALoDLM-8B), and the authors’ repository README (github.com/amazon-science/ALoDLM, mirror of the paper’s claims). Numbers below are the authors’ own, quoted as theirs.

What it is

A diffusion language model over a Qwen3 backbone, cut into a prelude / a shared recurrent core / a coda. At 8B the recurrent core is the middle 16 transformer layers, run up to 4 times. Each pass reads out predictions for the unresolved positions, commits the confident ones as discrete context, and keeps refining the latents of the rest until a learned halting criterion or the depth cap fires. Two inference knobs expose the trade: the entropy threshold τ (which tokens may commit) and the cumulative halt threshold q (when the inner loop stops).

The training objective is the part that matters for us: token-wise computation schedules are latent variables, learned under a conditional negative ELBO with sequence-level outcome credit for each halting decision, plus a geometric depth prior, relaxed depth-profile regularization, a first-pass control variate, weighted intermediate supervision and an auxiliary next-token loss. The gate and the backbone train jointly.

The numbers

scale	Qwen3 (its own AR backbone)	strongest DLM compared	ALoDLM
1.7B	63.8	61.0 (SDAR)	65.5
8B	78.5	75.1 (WeDLM)	80.3

Average over eleven benchmarks, single-token commitment per pass. So at both scales the looped model beats its own unlooped backbone by 1.7 and 1.8 points, and the best diffusion baseline by 4.5 and 5.2.

Efficiency (their B200, single stream): ~2.7x the throughput of vLLM-served Qwen3-8B on GSM8K at comparable or higher accuracy, and 133.5 GFLOPs/token at the 93.25% accuracy threshold against WeDLM’s 154.6 (-13.6%).

And the one result that makes the mechanism credible rather than a knob: numerical tokens learn a lower first-pass halting probability — the difficulty signal emerges from the objective, with no difficulty labels.

Read against last night’s φ

docs/research/2026-10-11-recurrence-equivalence.md fitted an iso-depth law and measured φ = 0.46: one recurrence buys 0.46 of a unique parameter, so a looped block is a worse use of a parameter than a distinct one at matched compute. These two results do not contradict each other, and the difference is the currency:

• iso-depth prices depth: recursion as extra capacity per parameter.
• ALoDLM prices allocation: recursion as difficulty-matched compute spent during generation, where the win shows up as tokens per second at a quality floor, and as a quality gain over the SAME backbone (their +1.8 at 8B), which is not what the φ law was measuring.

So the honest combination for our queue is not “recursion is cheap” and not “recursion is expensive” — it is “recursion buys quality only when its schedule is learned, and the schedule is what has to be measured”. Which is exactly the verdict metric the φ protocol change already asks for, now with a second paper saying the same thing from the other side.

What this changes in our queue, concretely

1. §3.3’s “adaptive depth is only an arm, neither --rand-depth nor MoR has been A/B’d against fixed depth” now has a current, working reference implementation — and a reference objective. Our ADR-0013 deletion (no halt head, honest CE, fixed depth won twice) was measured on this stack at our scale, so it stands until it loses an A/B; but a queue item that used to be “someday” is now “there is a published recipe with released weights and code”.
2. A new arm for AB-PROTOCOL, as a lane and not yet a run: gate the loop’s depth on the controller’s own confidence, trained through an allocation objective rather than a standalone ponder loss. The pieces already exist here — the controller’s sigmoid weights (w_attn/w_mem/w_ffn), per-iteration CE (L_Rec), the DSpark next-K head as a commitment target, and MoR’s per-slot ranking. What does not exist is the objective, and the objective is what made ALoDLM’s version work.
3. §3.6’s EBT item 2 (“halt when ΔE between iterations plateaus”) gets a competitor worth measuring against: a learned gate with outcome credit. Cheap first step is still the diagnostic — log the per-iteration ΔE next to a would-be gate signal, and check whether they agree.

What it does NOT license

• Our ADR-0013 is not overturned. Their win is at 1.7B/8B on a token-parallel diffusion backbone with a halting objective; ours was measured on a 9M byte-level AR loop with honest unweighted CE. Two different experiments.
• The diffusion part is not ours to adopt. §3.6 already records why (dormouse is AR; diffusion-EBM rejected), and the throughput win is inseparable from parallel token commitment — our byte stream commits one byte at a time.
• The costs are real and their limitations say so: longer time to first token, depth-specific KV caches, and an input-dependent throughput that can reverse the advantage on hard prompts. Our launch-bound step (§3.1’s 13.3% utilisation) is the opposite of their B200 profile, so no speed claim transfers either way.

Cost of the lane here, roughly

A halting head over the loop state (one small linear per iteration), an allocation term on the objective (the ponder/KL machinery ADR-0013 deleted, in a different shape), and a depth prior + intermediate supervision we already pay through L_Rec. The measurement plan is the one the protocol already demands: 3 seeds, one batch size, fixed depth control, 2k steps, and the depth curve from --eval-depths alongside BPB. Nothing here is shippable tonight, and nothing here says the deletion was wrong — it says the queue item now has a price and a reference.