The figure is built from the paper’s final-block deployable TTFT and greedy accuracies for RiM and direct-answer SFT. Table 1 numbers (TTFT + greedy accuracy)

The reported latency comparison is based on the paper’s latency table (tokens generated and wall-clock time per GSM8K question). Table 7 latency evidence

Core mechanism: fixed “memory blocks” as latent working memory

• Memory blocks are fixed sequences of special tokens appended to the question; the model processes the augmented sequence in a single forward pass. RiM memory-block definition + single-pass claim
• Training decouples internal computation from externalized intermediate text via a custom attention mask that supports multiple readouts while avoiding leakage from later ground-truth reasoning steps. Custom attention mask and leakage prevention
• Only special-token embeddings are updated (existing token embeddings are frozen) via LoRA adapters, to isolate learning to “use” memory blocks rather than rewriting pretrained token semantics. Frozen token embeddings + special tokens

Two-stage curriculum: “ground” then “refine”

• Stage 1: after each memory block, the model predicts the next reasoning step (teacher-forced), with dense supervision and an annealed weighting schedule across blocks. Stage 1 objective and supervision
• Stage 2: remove step-level targets; instead, readouts after each memory block refine/predict the final answer, matching the fixed-budget inference setting. Stage 2 answer refinement objective

(1) Memory blocks become input- and block-dependent latent workspaces

The paper reports a representation analysis where memory-block representations move along distinct training trajectories and spread out in PCA after training, consistent with learned, block-specific latent computation rather than fixed placeholders.

Skeptical note: this is still an indirect mechanism probe (representations ≠ causal intermediate computation proof). The paper also reports that within-block attention structure ablations (e.g., bidirectional attention inside a block) yield mixed results, suggesting inductive biases matter. Bidirectional within-block attention mixed results

(2) Performance improves at low inference latency

Main results compare RiM final-block deployable outputs to SFT w/o CoT, SFT w/ CoT, Coconut, and (contextually) DART. The paper also uses cross-validation-based checkpoint selection to reduce selection overfitting.

Skeptical note: DART comparisons use official numbers because code/reimplementation was unreliable; this reduces controllability (training budget, data, eval protocol). DART comparison is not fully controlled

(3) Robustness across inference-time memory budgets

The paper claims Stage 2 “converts” grounded blocks into a stable latent workflow such that accuracy remains stable across a range of block sizes and counts.

Skeptical note: robustness is demonstrated on GSM8K-derived evaluations; the paper does not show (in the provided text) whether this holds for different reasoning domains or longer/variable-length chain structure beyond the dataset’s supplied reasoning-step granularity. Benchmark scope for robustness experiments

A) Benchmark and objective mismatch risks

• The training curriculum is mathematically grounded with explicit reasoning-step supervision in Stage 1; this could bias the learned workspace toward step-aligned structure rather than general “latent workspaces” usable for broader domains. Stage 1 supervision source + math benchmark scope
• Comparisons to literature baselines are complicated by training/evaluation differences (notably DART, where code was unavailable). Non-controlled baseline comparison

B) Mechanistic inference uncertainty

• Representation analyses (PCA trajectories, cosine similarity drops, linear probe separability) provide supportive evidence, but they are not a direct causal test that memory blocks are strictly responsible for the reasoning improvements. Representation and probe-based mechanistic evidence
• Attention-mask design and token budgeting (M=2 per block, K=8 blocks; Stage 1 depth up to number of reasoning steps) could be doing much of the heavy lifting in the benchmark; generalizing these design choices to other datasets/tasks could require new tuning. Token budgeting and curriculum assumptions

C) Selection-bias control is good, but external validity remains unknown

The paper attempts to reduce selection overfitting in checkpoint selection (cross-validation protocol). That strengthens internal validity for reported numbers, but does not address whether the learned memory-workspace strategy generalizes to non-math reasoning formats, longer contexts, or tasks without step-aligned supervision. Checkpoint selection bias mitigation + evaluation scope

• Test out-of-domain latent workspace: evaluate RiM-style memory blocks on reasoning tasks where “step supervision” is either absent or structured differently from GSM8K. If RiM’s gains collapse without step-aligned traces, the “working memory for latent reasoning” interpretation would be narrower than claimed. Scope limitation in current evidence
• Do causal ablation of memory tokens: after Stage 2 training, systematically replace/mask memory blocks at inference and measure accuracy drop. If the improvements persist, “workspace” may be an over-interpretation. If they vanish, it becomes stronger mechanistic evidence. What is (and isn’t) tested causally
• Fairer baseline reimplementation: reimplement DART under comparable training budgets and the same evaluation protocol to reduce contextual comparison ambiguity. Baseline comparability uncertainty