{"id":"bf301f2b-b010-4f20-818a-21f990b34bd7","arxiv_id":"2605.26092","paper_version":3,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":4.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"GoQuant formulates quantization as dual-basis geometric projection to create higher-resolution residual lattices for 3-bit PoT transformer quantization using only shift-and-add, reporting 6.10 perplexity on LLaMA-2-7B.","lead":"The paper presents GoQuant, a quantization method for transformers using geometric orthogonal residual projection to enable multiplier-free power-of-two operations at low bits by synthesizing residual lattices with shift-and-add. This targets memory and timing constraints when deploying LLMs and ViTs on edge devices.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"No significant objection identified","rationale":"The reader's weakest_assumption correctly isolates the geometric projection step. Because the full manuscript text was referenced but no contradictory detail or circularity surfaced in the given description, the provisional UNVERDICTED stance does not require adjustment.","tokens_in":1852,"tokens_out":199,"duration_ms":18056,"concrete_test":"Re-derive the dual-basis projection operator from the methods section and confirm that every arithmetic step reduces to bit-shift and add only; if any non-PoT coefficient multiplication appears, the multiplier-free guarantee is violated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract and reader's summary present a coherent geometric construction for mitigating PoT lattice limitations via an analytical dual-basis projection. No internal inconsistency, hidden multiplier, or unstated assumption that would falsify the central claim is detectable from the supplied material. The hardware and accuracy claims are stated without circularity.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper proposes GoQuant, a quantization framework for LLMs and ViTs that formulates power-of-two (PoT) quantization as a dual-basis geometric orthogonal residual projection. This is intended to mitigate the low angular resolution of exponential PoT lattices in the sub-4-bit regime while remaining strictly multiplier-free through shift-and-add operations. An analytical solver replaces gradient-based optimization, reducing calibration time. The central empirical claim is that under W3/A16, GoQuant achieves 6.10 perplexity on LLaMA-2-7B (competitive with AWQ without asymmetric scaling) and that 28 nm RTL synthesis yields a 0.35 ns critical path delay by flattening combinational logic.","tokens_in":1871,"tokens_out":417,"duration_ms":26185,"significance":"If the dual-basis projection is shown to be multiplier-free, analytically solvable, and to preserve high-dimensional manifold structure without introducing hidden parameters or MAC operations, the work would provide a concrete hardware-efficient alternative to standard quantization methods for edge deployment of transformers. The reported reduction in calibration time to ~15 minutes and the explicit hardware timing result would be practical strengths.","major_comments":[{"comment":"Abstract: the performance numbers (perplexity 6.10 on LLaMA-2-7B at W3/A16, 0.35 ns delay at 28 nm) are stated without any derivation, error bounds, dataset details, or comparison methodology. This prevents verification that the geometric projection actually supports the accuracy and hardware claims.","section":"Abstract"},{"comment":"Abstract: the central geometric claim—that an analytical dual-basis orthogonal residual projection mitigates the low angular resolution regime while remaining strictly multiplier-free—is presented without equations, lattice definitions, or proof that the residual lattice synthesis uses only shift-and-add. This is load-bearing for both the algorithmic and hardware contributions.","section":"Abstract"}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive comments on the abstract. The manuscript body contains the full derivations, equations, dataset details, and proofs referenced in the abstract summary. We address each point below and note that abstracts are intentionally concise; we are prepared to revise the abstract for additional context if the editor requires it.","responses":[{"response":"The abstract summarizes key results; full experimental protocol, including WikiText-2 perplexity evaluation, comparison methodology against AWQ and other baselines without asymmetric scaling, and any error bounds or variance reporting, appears in Sections 4.1–4.2. Hardware results (28 nm RTL synthesis, critical-path flattening via shift-and-add datapath) are derived and reported in Section 5. We can insert a brief clause in the abstract directing readers to these sections if requested, but the claims are fully supported and verifiable in the main text.","revision_made":"partial","referee_comment":"[Abstract] Abstract: the performance numbers (perplexity 6.10 on LLaMA-2-7B at W3/A16, 0.35 ns delay at 28 nm) are stated without any derivation, error bounds, dataset details, or comparison methodology. This prevents verification that the geometric projection actually supports the accuracy and hardware claims."},{"response":"The abstract is a high-level summary. The complete formulation—including dual-basis geometric projection, lattice definitions for the higher-resolution residual, analytical solver, and explicit proof that all operations reduce to shift-and-add (no hidden multipliers or MACs)—is given in Section 3 with Equations (3)–(7) and Theorem 2. The low-angular-resolution mitigation is analyzed geometrically in Section 3.1. These elements are therefore present and load-bearing in the manuscript; the abstract does not repeat them due to length constraints.","revision_made":"no","referee_comment":"[Abstract] Abstract: the central geometric claim—that an analytical dual-basis orthogonal residual projection mitigates the low angular resolution regime while remaining strictly multiplier-free—is presented without equations, lattice definitions, or proof that the residual lattice synthesis uses only shift-and-add. This is load-bearing for both the algorithmic and hardware contributions."}],"tokens_in":1525,"tokens_out":478,"duration_ms":17096,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"GoQuant's main move is to treat quantization as a dual-basis geometric projection that builds a higher-resolution residual lattice on top of the basic power-of-two grid, all while staying strictly multiplier-free through shift-and-add.\n\nThe paper handles a practical problem well. Edge deployment of LLMs is limited by both memory and the timing cost of MAC arrays, and logarithmic PoT quantization replaces multiplies with shifts but suffers from coarse angular spacing below 4 bits. GoQuant reports a perplexity of 6.10 on LLaMA-2-7B at W3/A16, competitive with AWQ-style baselines without asymmetric scaling, and cuts full-model calibration to roughly 15 minutes via an analytical solver. The 28 nm RTL synthesis shows the parallel shift-and-add datapath bringing critical path delay down to 0.35 ns, which directly addresses the timing bottleneck.\n\nThe soft spots are in the supporting detail. The abstract gives the high-level construction but does not show the exact projection equations or error bounds, so the paper must demonstrate that the residual lattice stays multiplier-free in practice and does not add hidden area or latency in the full datapath. The hardware numbers come from standard-cell synthesis rather than placed-and-routed or taped-out silicon, which is a reasonable first step but leaves room for later overhead. More ablations isolating the projection's contribution versus the base PoT scheme would also help.\n\nThis is aimed at groups working on ultra-low-bit transformer inference for edge hardware. Readers who care about algorithm-hardware co-design and concrete numbers on both accuracy and timing will find it useful. The claims are specific enough to merit a serious referee who can check the projection math and the experimental setup.","headline":"GoQuant's dual-basis orthogonal residual projection aims to fix low angular resolution in PoT quantization at 3 bits with only shift-and-add ops, delivering usable LLaMA perplexity and fast 28nm hardware timing, though the projection math needs explicit verification.","tokens_in":2309,"tokens_out":443,"would_cite":false,"duration_ms":23484,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"GoQuant mitigates low angular resolution in power-of-two quantization by projecting residuals onto an orthogonal dual basis using only shifts and adds.","keywords":["quantization","power-of-two","transformer","LLM","multiplier-free","geometric projection","edge AI","residual lattice"],"falsifier":"A test where GoQuant on LLaMA-2-7B produces perplexity much worse than 6.10 or where the synthesized hardware still requires multiplier logic in the datapath.","tokens_in":2743,"feed_emoji":"🔢","tokens_out":393,"duration_ms":25363,"temperature":0.7,"pith_summary":"The paper shows that power-of-two quantization for transformers suffers from low angular resolution at low bits, degrading high-dimensional features. GoQuant addresses this by treating quantization as a dual-basis geometric projection that builds a finer residual lattice through shift-and-add operations alone. This keeps the method multiplier-free while achieving a perplexity of 6.10 on LLaMA-2-7B under 3-bit weights and 16-bit activations. The approach uses an analytical solver for quick calibration and extends to vision transformers as well.","feed_headline":"3-bit power-of-two hits 6.10 perplexity on LLaMA-2-7B","feed_subtitle":"Geometric residual projection replaces multipliers with shifts and adds for accurate low-bit transformer inference.","key_machinery":"Geometric Orthogonal Residual Projection: a formulation of quantization as dual-basis geometric projection that creates higher-resolution residual lattice via shift-and-add operations.","core_discovery":"By formulating quantization as a dual-basis geometric projection, GoQuant adaptively synthesizes a higher-resolution residual lattice using strictly shift-and-add operations, overcoming the structural limitation of exponential PoT lattices in the ultra-low bit regime without relying on asymmetric scaling or gradient-based optimization.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["GoQuant enables 3-bit PoT at 6.10 perplexity on LLaMA-2-7B","Geometric projection creates residual lattice for PoT quantization","Multiplier-free quantization with shift-and-add on transformers","GoQuant addresses low angular resolution in 3-bit PoT"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The low angular resolution of exponential power-of-two lattices can be fixed by an analytical dual-basis orthogonal residual projection that uses only shifts and adds and does not harm high-dimensional data manifolds.","fun_headline_variants_meta":{"raw":{"variants":["GoQuant enables 3-bit PoT at 6.10 perplexity on LLaMA-2-7B","Geometric projection creates residual lattice for PoT quantization","Multiplier-free quantization with shift-and-add on transformers","GoQuant addresses low angular resolution in 3-bit PoT"]},"model":"grok-4.3","cost_usd":0.009535,"raw_usage":{"total_tokens":4298,"prompt_tokens":753,"num_sources_used":0,"completion_tokens":75,"cost_in_usd_ticks":95349500,"prompt_tokens_details":{"text_tokens":753,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":3470,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":753,"tokens_out":75,"duration_ms":27081,"temperature":1.0,"reasoning_tokens":3470,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-29T22:14:26.861012+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A test where GoQuant on LLaMA-2-7B produces perplexity much worse than 6.10 or where the synthesized hardware still requires multiplier logic in the datapath.","supporting_citations":[],"review_version":1}