{"id":"f2d7eaa1-42be-40f9-a869-f7dcbd894a19","arxiv_id":"2607.25915","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Penelope localizes latent reasoning to a narrow decoder interval with a persistent boundary memory, matching latent-reasoning accuracy at lower inference cost.","lead":"A new training and inference scheme called Penelope lets a decoder-only language model do latent 'thinking' by repeatedly refining a small internal memory through a short block of layers, instead of re-running the whole network or writing out long reasoning traces. On three structured-reasoning benchmarks it matches the accuracy of similar latent-reasoning methods while cutting measured inference latency by roughly a third to a half.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1's headline efficiency comparison uses Coconut at a fixed K=8, not its validation-optimal K=3 shown in Figure 2; the claimed 46.9% latency reduction and accuracy parity are not established under a symmetric tuning rule.","rationale":"The paper's central claim is a practical accuracy–efficiency tradeoff. The most direct evidence is the Table 1 comparison against Coconut. That comparison is the linchpin: if the baseline is evaluated at a depth that is not its best, the reported latency reduction and accuracy parity are artifacts of the evaluation protocol. Figure 2 already shows the authors knew how to select Coconut's depth and found K=3, yet the main table uses K=8. The memory-capacity concern raised by the reader is a legitimate generality caveat, but it does not threaten the demonstrated result on the tested tasks; the baseline asymmetry threatens the headline numbers themselves. I therefore disagree with the reader's choice of weakest assumption. The concern does not invalidate the architecture—the placement control and K=0 diagnostic provide internal evidence—so the verdict should remain CONDITIONAL, with the condition now including a symmetric baseline tuning.","tokens_in":11744,"tokens_out":18483,"duration_ms":174878,"concrete_test":"Re-run Deep ListOps with a symmetric validation protocol: for each of the three latent-stage seeds, select the smallest best K for Coconut over {1,...,6} exactly as done for Penelope; report three-run EM and paired latency for Coconut at its selected K, and recompute Table 1 with Penelope at its selected K. If Coconut's tuned EM exceeds Penelope's by >2 points and the latency reduction falls below ~15%, the headline accuracy-parity and 46.9% latency-reduction claims fail. A minimal version is to add a Table 1 row for Coconut at K=3 with three-run statistics and the same synchronized timing protocol.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative evidence for the accuracy–efficiency tradeoff is Table 1(a), which reports Penelope at validation-selected K∈{5,4,2} versus Coconut at fixed K=8 (46 vs 160 decoder-layer applications; 99.82 vs 188.15 ms). However, §4.1 states that validation independently selects Penelope's depth while Coconut's K is fixed, and Figure 2 explicitly applies the 'same smallest-best validation rule to both recurrent paths,' finding Coconut's validation maximum at K=3 (55.50% EM) rather than K=8. At the matched validation-optimal operating points (Figure 2), the comparison is 46 vs 80 layer applications and 105.10 vs 123.28 ms—a 14.7% latency reduction, not 46.9%. Moreover, Coconut at K=3 is more accurate than at K=8 on the representative checkpoint (55.50 vs 52.79 EM), so the paper's 'competitive accuracy' claim may weaken when the baseline is tuned. The paper does not report three-run EM or latency for Coconut at K=3, leaving the central comparison asymmetric. This is not a claim about the architecture's internal consistency; it is a claim about the fairness of the headline evidence.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"Penelope introduces a latent-reasoning framework for decoder-only Transformers in which the lower decoder prefix is evaluated once to form a boundary memory, and a selected output-side decoder interval is then iterated to refine a fixed-size latent interface (8 memory anchors, 2 readout anchors) before a single answer-context pass. A progressive CoT-to-latent curriculum transfers visible reasoning into this recurrent path. Experiments on Deep ListOps, ProsQA, and PrOntoQA with Llama-3.2-1B, plus a Qwen3.5-0.8B control, compare Penelope against Visible CoT, Coconut, CODI, and a full-decoder recurrence reference. The paper reports competitive exact-match accuracy with reduced measured inference latency, and includes matched source checkpoints, controlled optimizer-update/data-exposure budgets, audited splits, paired latency measurements, an explicit pre-answer serial-depth accounting formula, and ablation/placement controls.","tokens_in":12062,"tokens_out":6599,"duration_ms":60752,"significance":"If the reported results hold, Penelope offers a practically useful way to allocate test-time compute in latent space without repeatedly executing the full decoder, which is a meaningful step beyond full-decoder latent-recurrence methods. The paper's strengths include the careful matched-source evaluation, the synchronized latency measurements, the explicit serial-depth accounting in Eq. (15), and the mechanism controls (K=0 diagnostic, transition-unit ablation, and placement analysis in Table 2). The central caveat is that the headline efficiency gain is computed under an asymmetric depth-selection protocol; the paper's own symmetric comparison in Figure 2 shows a much smaller, though still positive, latency reduction.","major_comments":[{"comment":"The headline latency reduction is not based on a symmetric tuning rule. Table 1(a) compares Penelope at validation-selected K=5 with Coconut at fixed K=8 (46 vs 160 sequential decoder-layer applications; 99.82 vs 188.15 ms). Figure 2, however, applies the same smallest-best validation rule to both recurrent paths and reports Coconut's validation optimum at K=3 (55.50% EM, 80 layer applications, 123.28 ms). At the matched operating points the latency reduction is 105.10 vs 123.28 ms (14.7%), not 46.9%, and the workload reduction is 42.5% rather than 71%. Please report three-run test EM and latency for Coconut at K=3, and either use that point in the primary table or explicitly justify why Coconut's K is fixed at 8 while Penelope's is validation-selected.","section":"Table 1(a), §4.1–4.2, Figure 2"},{"comment":"The same asymmetry appears on ProsQA and PrOntoQA: Coconut uses 6 latent steps (its native schedule), while Penelope selects the smallest validation-optimal depth from K∈{1,...,6}. No validation sweep or selection curve is reported for Coconut on these tasks. If the intended protocol is a matched comparison, Coconut's validation-optimal operating point should be reported; if the fixed values are deliberate choices from the original methods, that should be stated explicitly. This is load-bearing for the claimed cross-task efficiency advantage.","section":"§4.1, Table 1(b,c)"},{"comment":"The paper never varies the latent interface size (m=8 memory slots, q=2 readout anchors) nor measures how much information the boundary memory actually carries. The claim that 'latent refinement can be localized' assumes that the fixed-size interface is sufficient for the problem state. The placement control in Table 2 shows localization is sufficient at this capacity, but a capacity sweep (varying m and q) would be needed to support the broader claim that localized refinement remains accuracy-preserving beyond the specific benchmarks. Without such evidence, the generality claim should be tempered.","section":"§3.5, Eq. (2)"}],"minor_comments":[{"comment":"The sentence 'The principal advantage... a 46.9% reduction' conflicts with the symmetric comparison reported in Figure 2. Please clarify that this is a paired comparison against a non-validation-optimal Coconut configuration, and present the symmetric result as the primary evidence for the efficiency claim.","section":"§4.2"},{"comment":"Add a footnote explaining why Coconut's K values in Table 1 are not validation-selected, and cross-reference Figure 2 for the validation-matched comparison.","section":"Table 1 caption"},{"comment":"Please report test (not just validation) EM for Coconut at K=3 and for Penelope at its selected K, and state whether the K=3 point is included in the three-run replication in Table 3(b).","section":"Figure 2"},{"comment":"The accounting formula C_preans_Penelope(K)=L+(K+1)r is correct for the evaluated topology, but the derivation is terse. A one-line derivation using ℓ_s = L−r would help readers verify the count.","section":"§3.2, Eq. (15)"}],"recommendation":"major_revision","confidential_remarks":"The paper's central empirical claim is currently overstated because the headline comparison withholds validation tuning from the baseline. However, the symmetric comparison in Figure 2 still shows a positive, if more modest, efficiency gain, and the architectural idea is defensible. I recommend major revision rather than rejection: the authors should add the missing Coconut K=3 results, realign Table 1 or clearly label the asymmetric comparison, and add an interface-capacity analysis or soften the claims accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Penelope is a real architectural contribution, not a repackaging. Localizing latent recurrence to a narrow decoder interval, with a fixed boundary memory, recurrent readout states, and a CoT-to-latent curriculum, is new relative to the Coconut/CODI/LOTUS line. The experiments are unusually careful for this area: same source checkpoint, same data, same update budgets, audited splits, paired latency measurements. The placement control in Table 2 and the K=0 diagnostic are genuinely informative—they show that localization, not some layer-specific magic, is doing the work.\n\nThe main soft spot is the headline efficiency number. Table 1 puts Penelope at validation-selected K against Coconut at a fixed K=8. Their own Figure 2 applies the same smallest-best validation rule to both, and Coconut's optimum is K=3. At those matched points, the latency advantage falls from 46.9% to about 15% (105 vs 123 ms), and the layer count advantage from 71% to 42.5%. The paper does report this asymmetry in Figure 2, but the abstract and Table 1 lead with the more flattering comparison. They also don't give three-run EM or latency for Coconut at K=3, so the variance there is unknown. Still, the conclusion holds up qualitatively: Penelope is at worst comparable in accuracy and clearly cheaper in sequential decoder depth. Just don't quote the 46.9% number.\n\nTwo other soft spots. No code, data, or supplement is provided, so replication is not currently possible. And the memory capacity is fixed at 8 slots and 2 readout anchors with no ablation—we don't know how much headroom that interface has. The closest localized-recurrence baselines (LOTUS, Intra-Layer Recurrence, Training-Free Looped Transformers) are cited but not compared, so we can't tell how much of the benefit is specific to Penelope's interface.\n\nAll that said, this is a solid paper with honest limitations. It deserves serious peer review. The main asks should be: release code and data, make the tuning rule symmetric in the main tables, and add a memory-capacity sweep. I'd be happy to see it in a reading group, and I'd cite it if I worked on latent reasoning.","headline":"Genuine architectural contribution with careful experiments, but the headline latency gain is oversold; the matched comparison still supports a real efficiency win.","tokens_in":12613,"tokens_out":4086,"would_cite":true,"duration_ms":34406,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Penelope localizes latent reasoning to a narrow decoder interval, matching full-decoder accuracy while roughly halving inference latency on structured reasoning tasks.","keywords":["latent reasoning","localized recurrence","decoder-only transformers","chain-of-thought distillation","inference latency","structured reasoning","test-time compute","recurrent memory"],"falsifier":"Run Penelope on a structured-reasoning task whose minimal state size provably exceeds 8 values (e.g., a multi-counter or multi-hop graph task), sweep the number of memory anchors m from 1 to 32, and compare against a full-decoder latent baseline: if accuracy is substantially lower at m=8 and does not recover when m increases, then the fixed interface—and with it the localization claim—does not transfer to tasks requiring richer state.","tokens_in":11562,"feed_emoji":"🧠","tokens_out":7944,"duration_ms":65258,"temperature":0.7,"pith_summary":"This paper tries to establish that a pretrained decoder-only language model can do latent, hidden-state reasoning by repeatedly executing only a narrow slice of its layers rather than the whole network. Penelope builds a fixed problem-state memory with one pass through the lower layers, then loops a late 5-layer interval K times to refine that memory, and only then generates the answer. Across three structured-reasoning tasks, this localized loop matches the accuracy of full-decoder latent reasoning (e.g., 52.25% vs 52.79% exact match on Deep ListOps) while cutting sequential decoder-layer applications from 160 to 46 and measured latency from 188 ms to 100 ms. The broader point is that additional reasoning compute can be allocated cheaply and locally instead of scaling parameters or generating long visible traces.","feed_headline":"Penelope loops 5 of 16 layers to halve latent-reasoning latency","feed_subtitle":"Penelope keeps accuracy within 0.54 point of a full-decoder latent model while cutting decoder-layer passes from 160 to 46.","key_machinery":"The localized latent interface: a decomposition of the decoder into a lower prefix computed once, a recurrent interval F[ℓ_s:ℓ_e) of width r=5, and a final answer pass. The prefix produces a problem-conditioned boundary memory M₀ (m=8 anchors) and readout state R₀ (q=2 registers); then, K times, the interval is applied to the memory plus step- and group-embeddings, updating it via a time-modulated GRU with residual adapter and gated interpolation (M_t = M_{t-1} + 0.5α_t(fM_t − M_{t-1})), while a readout GRU maintains answer-accessible state. A prompt-side KV cache built once allows the loop to attend to fixed context without re-running the prefix; the cost formula C = L + (K+1)r makes the ma","core_discovery":"The central claim is that localization is computationally sufficient: a one-time lower-prefix pass creates a persistent boundary memory (8 latent slots and 2 readout registers), and iterative refinement of only a selected output-side interval (5 of 16 layers) carries the same reasoning accuracy as full-decoder recurrence, at a marginal cost of 5 instead of 16 decoder-layer applications per refinement step. Training uses a progressive curriculum that removes visible chain-of-thought tokens one at a time and adds latent refinements, culminating in answer-only generation. On Deep ListOps, ProsQA, and PrOntoQA, Penelope matches or nearly matches the full-decoder latent baseline (within 0.54, 1.6","pith_inferences":["The fixed m=8 interface is the most likely bottleneck: if a task requires more state than 8 anchors can encode, accuracy would saturate; an obvious extension is to measure accuracy vs m on tasks with larger state requirements.","The efficiency advantage depends on the one-time lower-prefix cost being amortized; for very long prompts or very small K, the relative benefit may shrink, a regime not analyzed in the paper.","Because the loop's marginal cost is only r layers, the method pairs naturally with adaptive-depth halting mechanisms, which the paper leaves open.","Since placement mattered little on Deep ListOps, an input-dependent learned interval is a plausible extension, though no evidence yet shows it would improve the tradeoff."],"forward_implications":["Additional latent reasoning depth can be bought at r/L of the full-decoder cost, making test-time compute scaling cheaper on existing decoder-only models.","A small fixed latent interface (8 slots) is sufficient to carry problem state for structured reasoning, suggesting memory anchors can act as a persistent reasoning scratchpad.","Visible chain-of-thought can be progressively distilled into the latent loop with negligible accuracy loss on these tasks, so long visible traces are not strictly necessary.","Localization is robust to where the recurrent interval is placed (early/middle/output-side within 0.19 points), so the method is not tied to a particular layer choice.","The main empirical win is a large latency reduction at comparable accuracy, not a new accuracy ceiling; additional steps are not monotonic and validation-selected depths are best."],"fun_headline_variants":["Penelope localizes latent loops to 5 layers, halving inference latency","5-layer latent recurrence matches full-decoder reasoning at half the cost","Latent reasoning with 5-layer loop keeps accuracy, cuts decoder passes","Penelope: 5 of 16 layers loop for efficient structured reasoning","Localized latent recurrence: same reasoning, fewer decoder-layer passes"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The fixed 8-slot latent memory (with 2 readout registers) is assumed to be expressive enough to hold and refine all problem-state information the tasks need; if a task's state cannot be compressed into 8 vectors, the localized loop will saturate and accuracy will fall.","fun_headline_variants_meta":{"raw":{"variants":["Penelope localizes latent loops to 5 layers, halving inference latency","5-layer latent recurrence matches full-decoder reasoning at half the cost","Latent reasoning with 5-layer loop keeps accuracy, cuts decoder passes","Penelope: 5 of 16 layers loop for efficient structured reasoning","Localized latent recurrence: same reasoning, fewer decoder-layer passes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000284,"raw_usage":{"total_tokens":1520,"prompt_tokens":764,"completion_tokens":756,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":508,"completion_tokens_details":{"reasoning_tokens":662}},"tokens_in":508,"tokens_out":756,"duration_ms":6513,"temperature":1.0,"reasoning_tokens":662,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T01:06:41.502859+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Penelope on a structured-reasoning task whose minimal state size provably exceeds 8 values (e.g., a multi-counter or multi-hop graph task), sweep the number of memory anchors m from 1 to 32, and compare against a full-decoder latent baseline: if accuracy is substantially lower at m=8 and does not recover when m increases, then the fixed interface—and with it the localization claim—does not transfer to tasks requiring richer state.","supporting_citations":[],"review_version":1}