{"id":"cb235734-25ef-42a3-90df-338cf2569bec","arxiv_id":"2506.19451","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"SemPA-Look groups tokens into packets using a leave-one-out residual semantic score and a fixed-width lookahead search, matching near-optimal ATS at linear text-encoding complexity.","lead":"This paper proposes SemPA-Look, a greedy lookahead algorithm that groups text tokens into packets so that losing any one packet damages the sentence's meaning as little as possible. It reports near-optimal semantic similarity at linear computational cost on simulated packet erasure channels.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Near-optimality vs exhaustive search is shown only for N=2 packets (K=8, M=4); no optimal baseline is reported for N≥3, so the headline 'comparable to exhaustive search at linear complexity' is unvalidated in the regime where the greedy lookahead actually matters.","rationale":"The reader's verdict is CONDITIONAL, and I agree with that disposition. My stress-test focuses the condition more sharply: the empirical evidence for the central 'comparable to exhaustive search' claim is confined to N=2, a degenerate case where the surrogate is exact and the search space has only 35 partitions. The paper's main novelty—greedy lookahead with a fixed-size candidate pool—only becomes nontrivial for N>=3, and there the paper provides no optimal baseline. A skeptical reader should therefore not accept the headline claim's generality until an exhaustive-search comparison for N>=3 is supplied. I do not see a reason to reject the paper: the complexity analysis (O(N*P*k) text-encoding steps) is sound, the RSS surrogate is well-motivated by the first-order ATS expansion, and the N=2 experiments are encouraging. But the current evidence is insufficient to support the claim in the regime where it matters most. Hence the verdict remains CONDITIONAL, with the added condition that near-optimality be demonstrated for N>=3 (e.g., K=12, M=4 or K=15, M=3) against true exhaustive search on a nontrivial caption sample.","tokens_in":19039,"tokens_out":11763,"duration_ms":118373,"concrete_test":"Run exhaustive search on a sample of MS-COCO captions with K=12, M=4 (N=3) and K=15, M=3 (N=5) under p=0.25, word tokenization; for each caption compute ATS for the SemPA-Look output (P=10, k=4) and for the true ATS-optimal group. If the mean ATS gap grows substantially beyond the N=2 gap reported in Fig. 5 (about 0.009) or any caption shows gap > 0.05, the 'comparable to exhaustive search' claim does not extend to N>=3. A useful second readout is to run SemPA-Look with k=N-1 and P=200 as a near-optimal reference on the same captions; if the k=4, P=10 version degrades relative to it as N grows, the linear-complexity configuration is not preserving near-optimality.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that SemPA-Look 'achieves high ATS and LPIPS scores comparable to exhaustive search' at linear complexity (Abstract; Section V). The only comparisons against exhaustive full search are Fig. 5 and Fig. 14, both with K=8, M=4, hence N=2 packets. For N=2, the packet group is determined by the first packet (the complement is forced), the lookahead has exactly one feasible future packet, and the RSS surrogate exactly matches the ATS objective up to constants (ATS(G) = (1-p)^2 + (1-p)p[psi(C)+psi(W\\C)] + p^2 phi(empty,W)). Thus the near-optimality at K=8 is a weak test of the algorithm. For the longer-message experiments (K up to 60, N up to 12 in Figs. 8-11), the paper compares only against GA, Random PA, or no packetization; no optimal or near-optimal reference is given. The greedy selection with fixed candidate pool P=10 and lookahead depth k=4 could suffer compounding errors as N grows, and no experiment rules this out. Without an N>=3 optimal-baseline comparison, the claim that linear-complexity SemPA-Look remains 'comparable to exhaustive search' is unsupported in the regime where exhaustive search is actually infeasible.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies packetization of text tokens for transmission over a packet erasure channel, aiming to maximize the average token similarity (ATS) between the original and reconstructed token sequences. It formulates an ATS maximization problem P1, replaces it with a surrogate problem P2 based on a residual semantic score ψ(C,W)=φ(W\\C,W), and proposes SemPA-Look, a greedy lookahead algorithm that at each level samples candidate packets and k random disjoint future packets, scores them by their average RSS, and selects the best. The paper claims near-optimal ATS and LPIPS relative to exhaustive search at linear complexity, with experiments on MS-COCO and WikiHow, including a text-to-image downstream evaluation.","tokens_in":19363,"tokens_out":7504,"duration_ms":76751,"significance":"If the central claim were established, the work would offer a practical, low-complexity packetizer for semantic token communication, with a transparent complexity metric (number of text-encoding steps) and a channel-aware objective. The paper also includes useful ablations of lookahead depth k, population size P, packet size M, and message length K, and it evaluates downstream AIGC quality with LPIPS. However, the current evidence does not support the headline claim in the N≥3 regime, and the surrogate/lookahead approximations lack formal guarantees and lack systematic validation against an optimal baseline exactly where the heuristic matters. The significance is therefore conditional on closing these gaps.","major_comments":[{"comment":"The only comparisons against Full Search are for K=8, M=4, hence N=2 packets. For N=2, with G={C,W\\C}, the ATS is (1−p)^2 φ(W,W) + (1−p)p[ψ(C,W)+ψ(W\\C,W)] + p^2 φ(∅,W), so the average RSS objective in P2 is exactly affine in ATS, and the lookahead step has only one feasible future packet. These experiments therefore cannot support the abstract's claim that SemPA-Look is 'comparable to exhaustive search' in the regime where the lookahead heuristic is nontrivial. No optimal or near-optimal reference is provided for the N≥3 experiments in Figs. 8–11, so the headline claim is unvalidated exactly where full search is infeasible.","section":"Section V, Figs. 5 and 14"},{"comment":"The RSS surrogate ψ(C,W)=φ(W\\C,W) and the lookahead average Ψ are introduced as faithful proxies for the contribution of packet C to ATS, but no approximation bound or consistency argument is given. The Taylor derivation in Section IV-B, Eqs. (21)–(25), relates the sum of all packet RSSs to ATS only in the low-p limit and only for the full group; it does not justify the greedy per-level maximization of (29) nor the use of k randomly sampled future packets. The choice of RSS over TSS is supported only by a single-sentence example (Table II) and Fig. 2. As a result, the core mechanism of the algorithm is a heuristic whose failure modes are not characterized.","section":"Section IV-A and IV-C, Eqs. (15), (29)"},{"comment":"The statement that P1 is NP-hard is asserted without a reduction. The enumeration count K!/((M!)^N N!) shows only that full search is expensive, not that the problem is NP-hard. Either provide a reduction from a known NP-hard problem (for example, from a clustering or partitioning problem with a suitable embedding φ) or explicitly soften the claim to 'computationally intractable by enumeration.'","section":"Section III-A"},{"comment":"RandomDisjoint(L_i, M, k) is specified for a fixed k, but when the leftover set L_i is too small to support k disjoint packets of size M, the sampling is infeasible. The caption of Fig. 8 acknowledges that the feasible step size decreases with level, but the pseudocode and the complexity formula do not include the cap k ← min(k, floor(|L_i|/M)). Without this cap, the algorithm as written is underspecified.","section":"Algorithm 1, lines 2–6, and Eq. (28)"},{"comment":"The reported complexity of SemPA-Look is (N−1)·P·(k+1), but Algorithm 1 runs N levels, so the correct count is N·P·(k+1); the numerical speedup in Fig. 12 appears to use the N·P·(k+1) count rather than the table's formula. In addition, the 'full search' complexity of 2^K in Table III is inconsistent with Section III-A, which gives the number of feasible groups as K!/((M!)^N N!) and 2^N encodings per group; the table's 2^K understates full search for N>2 (e.g., at K=14, M=2, N=7, 2^14 versus 135,135×128). The complexity comparison underlying the speedup claims should be corrected.","section":"Table III and Fig. 12"}],"minor_comments":[{"comment":"The sentence beginning 'Compared to these lookahead applications, rather N (theookahead searches...' is garbled and should be rewritten.","section":"Section I-A"},{"comment":"The paragraph 'For example, in Fig. 13...' is repeated verbatim; remove the duplicate.","section":"Section V, Fig. 13 discussion"},{"comment":"φ(·,·) is defined for encoded text x and y but subsequently used with set arguments such as F(G) and W; the paper should clarify how token order is handled when computing φ on unordered sets.","section":"Section II-B, Eq. (2)"},{"comment":"The text says candidate packets are sampled 'with replacement' while line 3 samples from L(ℓ); clarify whether duplicate candidate packets are allowed and how this relates to the statement about 'sampling with replacement for candidate tuples generation.'","section":"Algorithm 1"},{"comment":"The claimed linear complexity is in N for a fixed packet size M; since N=K/M, the dependence on K should be stated explicitly as linear in K only when M is fixed.","section":"Table III and Section IV-C"},{"comment":"Please fix 'unreplacement' (should be 'without replacement') in Section IV-C and 'between between' in the caption of Fig. 14.","section":"Minor typos"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a reasonable fit for eess.SP, but the main experimental weakness is the absence of any optimal-baseline comparison for N≥3, which is precisely the regime where the lookahead and RSS surrogate are supposed to add value. The authors should also clarify the relationship to their prior SemPA and SemPA-GBeam papers (refs. [19], [28]); the new lookahead contribution should be demonstrated in a setting where it is not degenerate. If the N≥3 gap can be addressed with a tractable optimal or near-optimal reference (or with a formal approximation argument), the paper could become publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a useful, incremental heuristic for semantic packet aggregation. The RSS surrogate is a sensible leave-one-out idea, and the lookahead sampling with linear per-message complexity is a reasonable way to avoid exponential search. The experiments on MS-COCO and WikiHow, including the image-generation quality check, give a fair picture of where the method helps: modest ATS gains over random grouping and over GA at a fraction of the compute. I believe the core algorithm works as described and the complexity table is honest about encoder calls. The soft spots are real but not fatal. The strongest one, which the stress-test note nails, is that the 'comparable to exhaustive search' result is shown only for K=8, M=4, i.e., N=2 packets. In that regime the RSS surrogate is exactly the ATS objective up to constants and the lookahead has one feasible future packet, so the near-optimality is almost built in. For N>=3, where exhaustive search is actually infeasible and the greedy lookahead matters, there is no optimal or near-optimal reference; the comparisons against GA and Random PA are informative but they do not support the abstract's claim about exhaustive search. The paper should either add a small N=3 or N=4 exhaustive baseline (maybe with shorter messages) or soften the claim. Second, P1 is asserted NP-hard without a reduction. That may be true, but the paper does not prove it, and the statement is not needed for the main contribution. Also, the RSS surrogate and random lookahead sampling have no approximation guarantees; the paper should state explicitly that these are heuristics, or at least stop implying otherwise. Minor issues: Algorithm 1 does not specify how k is capped when leftover tokens cannot support k disjoint packets (the text mentions this only in passing for K=24). Several figures lack error bars, and no code or full experimental configuration is released, which makes the exact ATS numbers hard to reproduce. The Taylor expansions in Section IV-B are fine as intuition but not rigorous bounds. On the citation pattern: the self-citation to SemPA and SemPA-GBeam is appropriate, and the related work is reasonable. Nothing here smells like padding. Overall, this is a solid subfield contribution that needs one substantive revision: validate or qualify the near-optimality claim for N>=3. It deserves a serious referee rather than a desk reject, and I would expect the paper to be acceptable after that gap is closed. For your reading group, it is a maybe - worth a look if you care about semantic communication or packetization heuristics, but not a must-read.","headline":"Incremental but honest algorithmic contribution; the near-optimality claim is only tested in the easy N=2 regime, so the paper needs a stronger baseline before the headline holds.","tokens_in":710,"tokens_out":1881,"would_cite":false,"duration_ms":29765,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A lookahead packetizer groups tokens to preserve meaning under erasures, matching brute-force quality at linear cost.","keywords":["token communication","semantic packet aggregation","average token similarity","residual semantic score","lookahead search","packet erasure channel","semantic communication","wireless AIGC"],"falsifier":"Enumerate every feasible packet group for short captions (e.g., K=8 tokens in M=4 packets), compute the exact ATS of each, and check whether SemPA-Look's chosen group lands in the top decile of the exact-ATS ranking; if a large fraction of captions miss this bar, the surrogate is not faithful to the objective it replaces.","tokens_in":18812,"feed_emoji":"📦","tokens_out":6341,"duration_ms":64618,"temperature":0.7,"pith_summary":"Token communication breaks down when a packet is lost, because a missing token can change the meaning of an entire sentence. This paper tries to establish that the combinatorial job of grouping tokens into erasure-protected packets can be done almost as well as exhaustive search in time that grows linearly with message length. The proposed method, SemPA-Look, scores each candidate packet by the similarity left behind when that packet is removed, then uses a few randomly sampled future packets to look ahead before committing to a grouping. On caption and image-generation tasks it reports average token similarity and perceptual image scores close to exhaustive search while cutting computational cost by up to 40 times relative to full search and 10 times relative to a genetic-algorithm baseline.","feed_headline":"Token packetizer matches exhaustive search at 40x lower cost","feed_subtitle":"Lookahead sampling groups tokens so a lost packet costs less meaning, at linear computation cost.","key_machinery":"The mechanism that carries the argument is the pair (RSS, lookahead search). RSS, $\\psi(C,W)=\\varphi(W\\setminus C,W)$, is a token-level surrogate for the message-level ATS: it requires one text-encoding step per packet rather than $2^N$ per packet group. The lookahead search then prevents greedy early choices by evaluating each candidate packet together with $k$ randomly sampled disjoint packets from the remaining tokens, using the average RSS $\\Psi(\\tilde{C}_i^{(\\ell)},W)$ of candidate plus lookahead packets as the selection score. Sampling candidates with replacement explores diversity; sampling lookahead packets without replacement respects the constraint that packets must form a disjoint partition. The algorithm proceeds level by level until all tokens are assigned, giving linear complexity in the number of packets.","core_discovery":"On the paper's own terms, the central discovery is that packet loss resilience in token communication can be cast as a per-packet scoring problem. The residual semantic score $\\psi(C,W)=\\varphi(W\\setminus C,W)$ measures how much of the original message's meaning survives when packet $C$ is erased, and maximizing the average of these scores over the packet group is shown empirically to track the true message-level average token similarity (ATS). Combining this surrogate with a level-wise lookahead search — sample $P$ candidate packets, sample $k$ disjoint future packets from the leftover tokens, average their RSS, pick the best candidate, and repeat — gives a packetization that is near-optimal in ATS and in LPIPS on reconstructed images, while needing only $(N-1)P(k+1)$ text-encoding steps instead of the exponential count of full search.","pith_inferences":["Editorial inference: Because both the search surrogate and the reported ATS use the same CLIP-based similarity function, the method's semantic fidelity is only as good as that encoder's notion of meaning; re-evaluating with an independent text similarity measure or human ratings would test whether the gains generalize.","Editorial inference: The same candidate-plus-lookahead scoring could be applied to other token-arrangement problems where a combinatorial search is prohibitive, such as ordering tokens under a fixed output budget or choosing which tokens to drop under truncation.","Editorial inference: The convergence of ATS as lookahead depth k grows suggests an adaptive rule that stops increasing k when the average RSS stabilizes, which could cut the constant factor in the linear complexity further."],"forward_implications":["If SemPA-Look is correct, a practical transmitter can packetize text with near-optimal semantic preservation under packet erasure without exhaustive search, making semantic token communication feasible for longer messages.","The reported 40x complexity reduction over full search and 10x over GA at comparable ATS means the gap between optimality and tractability is much smaller than the exponential worst case suggests.","RSS-based grouping dominates top-token and random grouping for outage probabilities below roughly 0.6, with an intermediate packet length (M=2 in the experiments) balancing packet loss against semantic loss.","The caption-to-image experiments imply that packetization choices directly shape downstream generative quality: preserving the RSS-relevant tokens keeps the generated image semantically aligned with the intended caption."],"supporting_citations":[{"why":"Prior greedy SemPA that this work extends; supplies the ATS-optimal packet grouping objective and demonstrates its high computational cost.","marker":"[19]"},{"why":"SemPA-GBeam, the genetic-beam baseline whose exponential complexity motivates the lookahead surrogate.","marker":"[28]"},{"why":"CLIP, the pre-trained encoder used by the similarity function that defines both RSS and ATS.","marker":"[31]"},{"why":"Lookahead search, the algorithmic inspiration for sampling future packets before committing to the current one.","marker":"[20]"},{"why":"Genetic algorithm, the linear-complexity baseline whose text-encoding count is compared against SemPA-Look.","marker":"[33]"},{"why":"MS-COCO dataset, source of the image-caption pairs used in the ATS and AIGC experiments.","marker":"[34]"},{"why":"Stable Diffusion, the model used at the receiver to generate images from reconstructed captions for LPIPS evaluation.","marker":"[36]"}],"fun_headline_variants":["Lookahead packetization matches exhaustive search at 40x speed","Linear-time token grouping with residual semantic score","Token loss resilience via lookahead search and residual scoring","Near-optimal token packets in linear time via lookahead","Semantic packet aggregation: 40x faster than exhaustive search"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument rests on the premise that the leftover-token similarity score, averaged over a few randomly sampled future packets, faithfully predicts which packet grouping will best preserve meaning under erasures; the paper gives no bound for this approximation.","fun_headline_variants_meta":{"raw":{"variants":["Lookahead packetization matches exhaustive search at 40x speed","Linear-time token grouping with residual semantic score","Token loss resilience via lookahead search and residual scoring","Near-optimal token packets in linear time via lookahead","Semantic packet aggregation: 40x faster than exhaustive search"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000656,"raw_usage":{"total_tokens":3055,"prompt_tokens":1050,"completion_tokens":2005,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":666,"completion_tokens_details":{"reasoning_tokens":1923}},"tokens_in":666,"tokens_out":2005,"duration_ms":15650,"temperature":1.0,"reasoning_tokens":1923,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:07:20.529974+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Enumerate every feasible packet group for short captions (e.g., K=8 tokens in M=4 packets), compute the exact ATS of each, and check whether SemPA-Look's chosen group lands in the top decile of the exact-ATS ranking; if a large fraction of captions miss this bar, the surrogate is not faithful to the objective it replaces.","supporting_citations":[{"cited_title":"Semantic Packet Aggregation and Repeated Transmission for Text-to-Image Generation","cited_arxiv_id":"2503.23734","evidence_quote":"Prior greedy SemPA that this work extends; supplies the ATS-optimal packet grouping objective and demonstrates its high computational cost."},{"cited_title":"Learning transferable visual models from natural language supervision,","cited_arxiv_id":null,"evidence_quote":"CLIP, the pre-trained encoder used by the similarity function that defines both RSS and ATS."},{"cited_title":"A review on genet ic algorithm: past, present, and future,","cited_arxiv_id":null,"evidence_quote":"Genetic algorithm, the linear-complexity baseline whose text-encoding count is compared against SemPA-Look."},{"cited_title":"Microsoft COCO: Common objects in context,","cited_arxiv_id":null,"evidence_quote":"MS-COCO dataset, source of the image-caption pairs used in the ATS and AIGC experiments."},{"cited_title":"High- resolution image synthesis with latent diffusion models,","cited_arxiv_id":null,"evidence_quote":"Stable Diffusion, the model used at the receiver to generate images from reconstructed captions for LPIPS evaluation."}],"review_version":1}