{"id":"aa00236e-8e22-45c7-bc6e-098c4da171c9","arxiv_id":"2607.08221","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"A tokenizer-free pixel embedding, position encoding, and 256-way head let frozen LLMs act as portable entropy models for lossless RGB compression across model families.","lead":"LUMI compresses RGB images losslessly by feeding raw pixels into frozen LLMs through a small learned embedding interface instead of text tokens. This removes tokenizer dependence and works across LLaMA, Qwen, and Gemma while matching or beating prior LLM compressors on natural, medical, and remote-sensing images.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Competitive rates rest on a weak baseline set and independent 16×16 patches; gains over classical codecs are small and may not survive stronger learned codecs or inter-patch context.","rationale":"The reader correctly flags independent 16×16 patches as the weakest modeling assumption and keeps the verdict CONDITIONAL for missing artifacts. That is real but secondary: even with perfect patches the paper’s performance claim is under-supported relative to the broader lossless literature it cites. The tokenizer-agnostic interface and multi-backbone consistency (Tables II–IV, ablations V–X) are solid engineering evidence and do not collapse under this concern; they simply do not yet establish that frozen-LLM pixel-space adaptation is competitive with specialized codecs. Hence I keep CONDITIONAL (same direction as the reader) but shift the load-bearing soft spot from “patches alone” to “patches + incomplete baseline set.” Agreement is partial because the reader’s weakest_assumption is necessary but not the single most decisive threat to the strongest_claim as written. No code/error bars remain secondary reproducibility issues, not the primary scientific soft spot.","tokens_in":16337,"tokens_out":735,"duration_ms":8673,"concrete_test":"Re-evaluate the same Kodak/BRACS/BED4RS test splits with at least one stronger public learned lossless baseline (e.g., a recent hierarchical residual or bit-plane model from the §II-A citations) under identical RGB patching, and report BPP side-by-side with LUMIQwen. If LUMI is worse by >0.3 BPP on two of three domains, the “competitive rates” claim should be narrowed to “competitive among frozen-LLM interfaces.”","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim (Tables III–IV) is that LUMI with frozen multi-family LLMs is competitive and more robust than pixel-as-text LLM baselines. That claim is only weakly load-bearing for the broader framing of “competitive lossless RGB compression.” In-domain Kodak rates (8.56–8.70 BPP) sit near JPEG-XL (8.95) and DLPR (8.74) and only slightly better than LLaMA (P2-LLM LoRA) at 8.60; BRACS/BED4RS gains are clearer but still modest. The paper does not compare against stronger modern learned lossless codecs (e.g., hierarchical residual / flow / bit-plane models cited in §II-A but not re-run). Combined with independent non-overlapping 16×16 patches (T=768, no inter-patch context; §III-G, Limitations), the frozen LLM is never asked to model long-range image statistics that classical and specialized neural codecs exploit. Thus “competitive” is demonstrated mainly against JPEG-XL, DLPR, and a tokenizer-bound LLM baseline under a deliberately restricted context regime, not against the best available entropy models. If those missing comparisons reverse the ranking, the engineering contribution remains (tokenizer-agnostic interface) but the headline performance claim weakens.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.5","summary":"LUMI proposes a tokenizer-agnostic interface for lossless RGB image compression that attaches lightweight external modules (PixEmb intensity–channel embedding, gated row–column intra-patch position encoding, soft prefix, and a 256-way pixel head) to frozen decoder-only LLMs (LLaMA, Qwen, Gemma). Pixel values are mapped into continuous embedding space rather than textual numeric tokens, so the same pipeline avoids family-specific tokenizer fragmentation and predicts a native 256-symbol alphabet for arithmetic coding. Only the external adapters are trained. Experiments on Kodak, BRACS, and BED4RS report in-domain BPP competitive with JPEG-XL and DLPR and with a LoRA-adapted P2-LLM baseline, leave-one-domain-out generalization, ablations of PixEmb/INP/prompting, and model/data scaling on Qwen3.","tokens_in":16732,"tokens_out":1081,"duration_ms":18384,"significance":"The main contribution is a clean formulation of LLM-based lossless image coding as pixel-space adaptation of frozen foundation models rather than tokenizer-bound language-symbol modeling. Demonstrating a single external interface that works across three tokenizer families, with only small trainable modules and held-out BPP measured under arithmetic coding, is a useful engineering and conceptual result for the compression and foundation-model communities. Tokenizer-fragmentation evidence (Table II), embedding visualizations, and consistent multi-backbone behavior strengthen the portability claim. The work is appropriately scoped as an interface study; it does not claim to redefine the state of specialized neural lossless codecs.","major_comments":[{"comment":"Tables III–IV and §IV-C frame LUMI as achieving “competitive” rates, but the non-LLM comparison set is limited to JPEG-XL and DLPR. Section II-A cites stronger modern learned lossless methods (hierarchical residual, invertible flows, bit-plane models) that are not re-run under the same patch protocol. Without those anchors, the headline performance claim is only weakly supported relative to the broader learned-codec literature, even though the gain over tokenizer-based P2-LLM baselines is clear.","section":null},{"comment":"Section III-G and the Limitations section state that non-overlapping 16×16 patches are compressed independently (T=768, no inter-patch context). This design choice is load-bearing for the claim that frozen LLMs act as effective image entropy models: classical and specialized neural codecs exploit longer-range structure that LUMI never models. The paper should either quantify the cost of this restriction (e.g., vs. larger patches or a simple inter-patch conditioner) or more carefully qualify “competitive” as holding under a deliberately restricted context regime.","section":null},{"comment":"Table I reports only 4 training images for BRACS (9,767 patches). In-domain and leave-one-domain-out BRACS numbers in Tables III–IV and the scaling tables therefore rest on a very small image-level sample. Without image-level resampling, error bars, or multi-split results, the medical-domain and cross-domain robustness claims are statistically under-supported relative to Kodak and BED4RS.","section":null}],"minor_comments":[{"comment":"No standard errors or run-to-run variance are reported for any BPP table; even a few seeds on the adapter training would help assess stability of the small gains over JPEG-XL/DLPR.","section":null},{"comment":"Figure 2’s embedding visualization is qualitative; a simple neighborhood or channel-separation metric would make the PixEmb vs. tokenizer comparison more rigorous.","section":null},{"comment":"Equation (11) introduces a 7-D descriptor with polynomial and Fourier features; a short justification or ablation of the sinusoidal terms (beyond 4-D vs. 7-D channel identity) would clarify design choices.","section":null},{"comment":"Decoding latency is noted as a limitation but never quantified; a brief wall-clock comparison against JPEG-XL/DLPR on the same hardware would contextualize practicality.","section":null},{"comment":"Minor notation: β in Eq. (14) is called a “learnable scalar gate” but its learned range or initialization is not reported; likewise soft-prefix length P=16 is fixed without sensitivity analysis.","section":null},{"comment":"Typographical consistency: “P 2-LLM” / “P2-LLM” spacing and “F . Arithmetic Coding” section heading spacing should be cleaned for camera-ready.","section":null}],"recommendation":"minor_revision","confidential_remarks":"Fit for a solid systems/methods venue in CV or compression. The tokenizer-agnostic interface is the real contribution; the performance framing is slightly oversold relative to the baseline set. I would not block on missing SOTA neural codecs if the authors rephrase “competitive” and add the statistical caveats on BRACS, but editors who prioritize absolute rate tables may want a major-revision bar. No integrity concerns."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The real contribution here is the interface, not a new compression theory. LUMI stops forcing RGB values through text tokenizers, maps intensity + channel + axial position into the frozen LLM embedding space, and predicts with a native 256-way head. That is a clear, portable design. They show it works on LLaMA, Qwen, and Gemma without touching backbone weights, and the tokenizer-fragmentation table (\"255\" is one token vs three) makes the motivation concrete.\n\nWhat they do well: the ablations are coherent. 7-D PixEmb beats 4-D; INP helps; soft prefix + task prompt interact sensibly; leave-one-domain-out and Kodak-trained transfer to BRACS/BED4RS support the \"reusable interface\" claim better than pure in-domain numbers. Scaling with Qwen size and BRACS patch count is honest and positive. Math is standard NLL + arithmetic coding; no circularity. Citations cover classical codecs, learned lossless work, and the P2-LLM line they improve on.\n\nSoft spots, in proportion: \"competitive\" is real but modest. Kodak sits near JPEG-XL and DLPR and only slightly ahead of LoRA P2-LLM; BRACS/BED4RS gains are clearer. They do not re-run stronger hierarchical residual/flow/bit-plane codecs they cite in related work, and they model non-overlapping 16x16 patches with no inter-patch context (they flag this themselves). So the frozen LLM never has to model long-range image structure that specialized codecs exploit. Medical training is tiny (4 images). No error bars, no code. Those limit how hard you can lean on the absolute rate claims; they do not erase the interface result.\n\nThis is for people working on LLM adaptation, reprogramming, or portable entropy models who care about multi-family portability more than beating the absolute SOTA codec. The central engineering claim holds under the regime they actually test. I would send it to peer review; a referee can demand stronger baselines and artifacts without the paper being empty. Worth reading if that interface problem is on your desk.","headline":"Clean tokenizer-free interface that makes frozen multi-family LLMs usable as pixel entropy models; rates are competitive mainly against JPEG-XL/DLPR and P2-LLM under independent 16x16 patches.","tokens_in":17321,"tokens_out":546,"would_cite":true,"duration_ms":6194,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Lossless image compression can be done by adapting frozen language models in pixel space, without using their text tokenizers.","keywords":["lossless image compression","large language models","entropy modeling","frozen foundation models","tokenizer-agnostic","pixel embedding","arithmetic coding"],"falsifier":"Train and evaluate the same external modules on identical data and splits, but with full inter-patch context or larger overlapping tiles; if rates do not improve over the independent-patch LUMI baseline, the sufficiency claim holds; if they drop sharply below classical codecs, the assumption fails.","tokens_in":17262,"feed_emoji":"🖼️","tokens_out":589,"duration_ms":6480,"temperature":0.7,"pith_summary":"This paper argues that the usual way of squeezing images through language models—turning pixel numbers into text and reading probabilities from the vocabulary—is the wrong interface. Tokenizers split the same number differently across model families, so the coding events themselves become family-specific. LUMI instead maps raw intensity, color channel, and local position straight into the continuous embedding space of a frozen decoder-only model, then predicts a clean 256-way distribution over pixel values for arithmetic coding. Only a small external adapter (pixel embedding, position tables, soft prefix, and output head) is trained. Across natural, medical, and remote-sensing images, and with LLaMA, Qwen, and Gemma backbones, the method matches or beats tokenizer-based LLM baselines and classical codecs while remaining portable across tokenizers. The broader claim is that foundation models can serve as reusable entropy engines once image symbols are presented in their native alphabet rather than as language tokens.","feed_headline":"Skip the tokenizer: freeze the LLM, compress pixels directly","feed_subtitle":"A small pixel adapter turns LLaMA, Qwen, and Gemma into portable lossless image codecs.","key_machinery":"The tokenizer-free pixel interface (PixEmb + intra-patch position encoding + 256-way head): a 7-D intensity/channel descriptor projected into LLM embedding space, plus row-column position codes, feeding a frozen decoder that emits exact pixel probabilities for arithmetic coding.","core_discovery":"LLM-based lossless RGB compression is best formulated as pixel-space adaptation of frozen foundation models: map intensity, channel, and intra-patch position into continuous embeddings, keep the backbone fixed, and decode with a dedicated 256-way head, rather than representing pixels as tokenizer-dependent text.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Freeze the LLM backbone: compress RGB via pixel embeddings not tokens","Tokenizer-agnostic LUMI maps pixels into frozen LLM embedding space","Pixel adapters turn LLaMA Qwen Gemma into portable lossless codecs","Lossless images by pixel-space adaptation of fixed foundation models","Skip tokenizers: 256-way head decodes intensity from frozen LLMs"],"cache_read_input_tokens":128,"weakest_assumption_plain":"That modeling each 16-by-16 patch independently with a frozen language model already captures enough local statistics for competitive compression rates.","fun_headline_variants_meta":{"raw":{"variants":["Freeze the LLM backbone: compress RGB via pixel embeddings not tokens","Tokenizer-agnostic LUMI maps pixels into frozen LLM embedding space","Pixel adapters turn LLaMA Qwen Gemma into portable lossless codecs","Lossless images by pixel-space adaptation of fixed foundation models","Skip tokenizers: 256-way head decodes intensity from frozen LLMs"]},"model":"grok-4.5","effort":"low","cost_usd":0.004766,"raw_usage":{"total_tokens":1393,"prompt_tokens":802,"num_sources_used":0,"completion_tokens":95,"cost_in_usd_ticks":47660000,"prompt_tokens_details":{"text_tokens":802,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":496,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":802,"tokens_out":95,"duration_ms":5208,"temperature":1.0,"reasoning_tokens":496,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T11:04:40.572204+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train and evaluate the same external modules on identical data and splits, but with full inter-patch context or larger overlapping tiles; if rates do not improve over the independent-patch LUMI baseline, the sufficiency claim holds; if they drop sharply below classical codecs, the assumption fails.","supporting_citations":[],"review_version":1}