Pith. sign in

REVIEW 1 major objections 70 references

Spherical KV stores keys as angles and radii to compute attention logits directly and allocates retention by rate-distortion utility under a fixed budget.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.3

2026-06-30 21:25 UTC pith:A3I55DDT

load-bearing objection Spherical KV pairs angle codes with rate-distortion retention but supplies no math or tests for the core approximation. the 1 major comments →

arxiv 2605.18856 v3 pith:A3I55DDT submitted 2026-05-13 cs.LG cs.CLcs.ITmath.IT

SPHERICAL KV: Angle-Domain Attention and Rate-Distortion Retention for Efficient Long-Context Inference

Anay Chauhan , Gurucharan Marthi Krishna Kumar , Arion Das , Amit Dhanda , Vinija Jain , Aman Chadha , Amitava Das This is my paper
classification cs.LG cs.CLcs.ITmath.IT
keywords KV cachelong-context inferenceattention geometryrate-distortionmemory efficiencydecode optimizationspherical parameterization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper seeks to ease the memory growth and HBM streaming limits that arise when KV caches expand with context length in language model inference. It does so by representing directional key information in compact spherical form so that attention can proceed without rebuilding full vectors, and by deciding what to keep or how precisely under a rate-distortion budget. The resulting pages stay block-local and fusion-friendly, directly cutting the traffic that dominates decode time. A reader would care because the approach aims at practical serving constraints rather than abstract compression alone.

Core claim

Spherical KV treats KV allocation as a rate-distortion problem grounded in attention geometry. Keys are stored as a scalar radius plus compact angle codes; attention logits are formed directly from these codes without reconstructing dense keys. Rate-Distortion Retention then chooses per-token, per-head keep or drop decisions and precision tiers inside a fixed budget, yielding tier-homogeneous pages with lightweight metadata that support coalesced reads.

What carries the argument

Angle-Domain Attention (ADA) that parameterizes keys spherically for direct logit computation, combined with Rate-Distortion Retention (RDR) that jointly optimizes retention and precision under a budget.

Load-bearing premise

The spherical angle codes and rate-distortion allocation preserve enough attention fidelity that downstream model accuracy stays acceptable.

What would settle it

Measure perplexity or needle-in-haystack accuracy on a long-context benchmark when Spherical KV is enabled versus a dense baseline; a clear drop beyond a few percent would falsify the fidelity claim.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • KV residency shrinks while the decode path remains paged, block-local, and fusion-friendly.
  • HBM traffic falls because dense keys are never reconstructed in the hot loop.
  • Tier-homogeneous pages with light metadata enable efficient coalesced memory reads.
  • The method supplies a deployment-oriented way to trade residency against compute without altering attention structure.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same spherical representation might pair with existing quantization or eviction schemes for additive savings.
  • Utility estimates inside RDR could be made task-aware if future utility signals become available at serving time.
  • The geometry assumption may extend to other attention variants if directional information dominates in those models as well.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

1 major / 0 minor

Summary. The manuscript proposes Spherical KV, a long-context inference technique that represents keys via a spherical parameterization (scalar radius plus compact angle codes) in Angle-Domain Attention (ADA) so that attention logits can be computed directly from the codes without reconstructing dense key vectors, paired with Rate-Distortion Retention (RDR) that jointly selects per-token/head keep/drop decisions and precision tiers under a fixed budget to produce tier-homogeneous pages; the combined method is claimed to reduce KV residency and HBM traffic while preserving decode efficiency and model accuracy.

Significance. If the geometric approximation in ADA maintains sufficient fidelity to the original attention scores, the approach could meaningfully alleviate the memory-bandwidth bottleneck in long-context serving by enabling paged, block-local, fusion-friendly decode paths; the rate-distortion framing supplies a principled allocation mechanism that existing eviction or quantization methods lack.

major comments (1)
  1. [Abstract] Abstract: the claim that attention logits are computed 'directly from these codes' without key reconstruction is load-bearing for both the efficiency and accuracy assertions, yet the manuscript supplies neither the explicit mapping from compact angle codes to approximated unit vectors, nor an error analysis or bound on how angle compression and quantization distort the geometry relative to the original Q·K dot products. Without this, it is impossible to verify whether per-head approximation error compounds across layers sufficiently to alter attention distributions and downstream outputs.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for highlighting the need for explicit details on the Angle-Domain Attention mapping and error analysis. This is a valid observation on the current manuscript draft. We will strengthen the presentation in revision while preserving the core claims.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that attention logits are computed 'directly from these codes' without key reconstruction is load-bearing for both the efficiency and accuracy assertions, yet the manuscript supplies neither the explicit mapping from compact angle codes to approximated unit vectors, nor an error analysis or bound on how angle compression and quantization distort the geometry relative to the original Q·K dot products. Without this, it is impossible to verify whether per-head approximation error compounds across layers sufficiently to alter attention distributions and downstream outputs.

    Authors: We agree the manuscript would be improved by including the explicit mapping and supporting analysis. In the revised version we will add: (1) the precise decoding function that maps the compact angle codes plus radius to an approximated unit vector via quantized spherical coordinates converted to Cartesian form, enabling direct logit computation as q · (r · û(code)); (2) a quantization-error bound derived from the angular granularity together with empirical per-layer logit distortion measurements on representative long-context traces showing that the approximation error remains small enough not to materially change attention distributions or downstream accuracy. These additions will be placed in the ADA section and referenced from the abstract. revision: yes

Circularity Check

0 steps flagged

No circularity in derivation chain

full rationale

The provided abstract and text introduce ADA and RDR as engineering components for KV cache management using spherical parameterization and rate-distortion allocation. No equations, fitted parameters called predictions, self-citations as load-bearing premises, or renamings of known results are present. The claims rest on the proposed mechanisms themselves without reducing to self-defined inputs or prior author work by construction. This is a standard non-finding for a methods paper whose central contribution is the design rather than a closed mathematical derivation.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no explicit free parameters, axioms, or invented entities; all such elements remain unknown.

pith-pipeline@v0.9.1-grok · 5809 in / 956 out tokens · 22462 ms · 2026-06-30T21:25:10.496127+00:00 · methodology

0 comments
read the original abstract

Long-context inference is increasingly constrained by the KV cache: resident memory grows with context length, and decoding becomes limited by repeated High Bandwidth Memory (HBM) streaming rather than arithmetic. Existing methods such as eviction, windowing, quantization, and offloading reduce footprint, but often leave the critical-path bottleneck only partially addressed, especially when compressed states must still be reconstructed into dense vectors during decoding. We present Spherical KV, a long-context inference method that treats KV allocation as a rate-distortion problem grounded in attention geometry for efficient decoding. The method is built on two ideas: (i) represent directional information cheaply in the decode hot loop, and (ii) allocate retention and precision according to estimated future utility. Its first component, Angle-Domain Attention (ADA), stores keys in a spherical parameterization consisting of a scalar radius and compact angle codes, and computes attention logits directly from these codes without reconstructing dense keys. This preserves a paged, block-local, fusion-friendly decode path and directly targets HBM traffic in realistic serving settings. Its second component, Rate-Distortion Retention (RDR), jointly chooses keep/drop decisions and precision tiers per token and head under a fixed budget, producing tier-homogeneous pages with lightweight metadata and coalesced reads. Together, ADA and RDR provide a deployment-oriented mechanism for reducing KV residency while preserving decode efficiency.

Figures

Figures reproduced from arXiv: 2605.18856 by Aman Chadha, Amitava Das, Amit Dhanda, Anay Chauhan, Arion Das, Gurucharan Marthi Krishna Kumar, Vinija Jain.

Figure 1
Figure 1. Figure 1: Spherical KV Architecture. Dense keys from the incoming layer are mapped to a spherical param￾eterization (radius r and angles ϕ). A learned rate–distortion retention policy π enforces a strict memory budget B by predicting (i) a keep/drop decision zi ∈ {0, 1} and (ii) a per-item precision (rate allocation) bi . Joint quantization and filtering produce a sparse, quantized spherical stream that is stored in… view at source ↗
Figure 1
Figure 1. Figure 1: Spherical KV. Dense keys from the incoming layer are mapped to a spherical parameterization with radius r and angles ϕ. A learned rate–distortion retention policy π enforces a strict memory budget B by predicting (i) a keep/drop decision zi ∈ {0, 1} and (ii) a per-item precision tier bi . Joint quantization and filtering produce a sparse, quantized spherical stream that is stored in a compressed KV cache i… view at source ↗
Figure 2
Figure 2. Figure 2: Angle-Domain Attention in Practice (Paged KV; kernel-realized, no reconstruction). This diagram shows how Angle-Domain Attention fits into a standard paged KV-cache decode loop without paying a hidden densification tax. Top (prefill): for each token, we write a paged/contiguous K-cache page that stores compact spherical K-codes—a scalar rk (radius) and c θ k (packed angle codes), plus lightweight tier/flag… view at source ↗
Figure 2
Figure 2. Figure 2: Angle-Domain Attention in Practice (paged KV; kernel-realized, no reconstruction). ADA integrates into a paged KV decode loop without densification. (i) Prefill: per token, we write a paged K-cache page stor￾ing compact spherical key codes—radius rk, packed angular codes c θ k , and tier/flags metadata in a header￾guided, kernel layout. In parallel, we write a V-cache page in FP16/BF16, preserving paged KV… view at source ↗
Figure 3
Figure 3. Figure 3: Rate–Distortion Retention: joint keep/drop + tiering under a strict KV budget. We treat KV residency as a constrained rate–distortion allocation problem: under a fixed byte budget, the controller spends preci￾sion on states that are likely to matter later and removes low-utility states first. (A) RD controller. Inputs in￾clude token/head/layer context plus a cost model; an RD score (e.g., predicted ∆loss v… view at source ↗
Figure 3
Figure 3. Figure 3: Rate–Distortion Retention: joint keep/drop + tiering under a strict KV budget. KV residency is treated as a constrained rate–distortion allocation problem: under a fixed budget, the controller spends pre￾cision on high-impact states and removes low-utility states first. (A) RD controller. Given token/head/layer context and a cost model, the controller assigns each state to drop or keep with a precision tie… view at source ↗
Figure 4
Figure 4. Figure 4: Iso-quality Pareto frontiers for memory-bounded decoding. Each panel plots decode throughput (tok/s; higher is better) versus effective KV budget bKV (bytes/token; lower is better) under paged/ragged serving, across three models and context lengths L ∈ {8K, 32K, 128K}. Let Q be the quality score (higher is better; defined in § [X]) and let Q⋆ dense denote the best Dense-KV quality in the panel. We enforce … view at source ↗
Figure 4
Figure 4. Figure 4: Iso-quality Pareto frontiers for memory-bounded decoding. Each panel plots decode throughput (tok/s; higher is better) versus effective KV budget bKV (bytes/token; lower is better) under paged/ragged serving, across three models and context lengths L ∈ {8K, 32K, 128K}. Let Q be the quality score (higher is better; defined in § [X]) and let Q⋆ dense denote the best Dense-KV quality in the panel. We enforce … view at source ↗
Figure 5
Figure 5. Figure 5: Ablations A0–A5 at extreme context (L=128K): frontier shift + mechanism evidence under iso￾quality (∆ = 0.8). Top row (A0–A3, three LLMs): For each model (columns), we plot decode throughput s (tok/s; higher is better) versus effective resident KV budget bKV (bytes/token; lower is better) measured on a paged/ragged KV substrate. Let Q⋆ dense be the best Dense KV quality for that model; we retain operating … view at source ↗
Figure 5
Figure 5. Figure 5: Ablations A0–A5 at extreme context (L=128K): frontier shift + mechanism evidence under iso￾quality (∆ = 0.8). Top row (A0–A3, three LLMs): For each model (columns), we plot decode throughput s (tok/s; higher is better) versus effective resident KV budget bKV (bytes/token; lower is better) on a paged/ragged KV substrate. Let Q⋆ dense be the best Dense KV quality for that model; we retain points satisfying Q… view at source ↗
Figure 6
Figure 6. Figure 6: Stability phase diagram and bounded gating behavior. (A) Phase diagram (policy map). The x-axis is angular precision b θ (equivalently, the effective KV budget allocated to angular codes), and the y-axis is a brit￾tleness proxy Bt that increases when attention becomes sensitive to small logit perturbations (e.g., inverse￾margin or a normalized query-norm proxy. The plane separates three regimes: safe compr… view at source ↗
Figure 6
Figure 6. Figure 6: Spherical KV Architecture. Dense keys from the incoming layer are mapped to a spherical param￾eterization (radius r and angles ϕ). A learned rate–distortion retention policy π enforces a strict memory budget B by predicting (i) a keep/drop decision zi ∈ {0, 1} and (ii) a per-item precision (rate allocation) bi . Joint quantization and filtering produce a sparse, quantized spherical stream that is stored in… view at source ↗
Figure 7
Figure 7. Figure 7: Angle-Domain Attention in Practice (Paged KV; kernel-realized, no reconstruction). This diagram shows how Angle-Domain Attention fits into a standard paged KV-cache decode loop without paying a hidden densification tax. Top (prefill): for each token, we write a paged/contiguous K-cache page that stores compact spherical K-codes—a scalar rk (radius) and c θ k (packed angle codes), plus lightweight tier/flag… view at source ↗
Figure 8
Figure 8. Figure 8: Rate–Distortion Retention: joint keep/drop + tiering under a strict KV budget. We treat KV residency as a constrained rate–distortion allocation problem: under a fixed byte budget, the controller spends preci￾sion on states that are likely to matter later and removes low-utility states first. (A) RD controller. Inputs in￾clude token/head/layer context plus a cost model; an RD score (e.g., predicted ∆loss v… view at source ↗
Figure 9
Figure 9. Figure 9: Iso-quality Pareto frontiers for memory-bounded decoding. Each panel plots decode throughput (tok/s; higher is better) versus effective KV budget bKV (bytes/token; lower is better) under paged/ragged serving, across three models and context lengths L ∈ {8K, 32K, 128K}. Let Q be the quality score (higher is better; defined in § [X]) and let Q⋆ dense denote the best Dense-KV quality in the panel. We enforce … view at source ↗
Figure 10
Figure 10. Figure 10: Kernel-realized efficiency witness under paged KV serving. (A) HBM bytes/token versus effec￾tive resident KV budget bKV. A representation-level improvement is only deployment-meaningful if lower resident bytes are accompanied by lower HBM traffic in the decode path. Spherical KV moves down-left relative to Dense KV and reconstruction-heavy baselines, showing that its memory savings survive the ker￾nel pat… view at source ↗
Figure 11
Figure 11. Figure 11: Ablations A0–A5 at extreme context (L=128K): frontier shift + mechanism evidence under iso￾quality (∆ = 0.8). Top row (A0–A3, three LLMs): For each model (columns), we plot decode throughput s (tok/s; higher is better) versus effective resident KV budget bKV (bytes/token; lower is better) measured on a paged/ragged KV substrate. Let Q⋆ dense be the best Dense KV quality for that model; we retain operating… view at source ↗
Figure 12
Figure 12. Figure 12: Stability phase diagram and bounded gating behavior. (A) Phase diagram (policy map). The x￾axis is angular precision b θ (equivalently, the effective KV budget allocated to angular codes), and the y-axis is a brittleness proxy Bt that increases when attention becomes sensitive to small logit perturbations (e.g., inverse-margin or a normalized query-norm proxy. The plane separates three regimes: safe compr… view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

70 extracted references · 1 canonical work pages

  1. [1]

    GEAR: An efficient KV cache compression recipe for near-lossless generative inference of LLM.arXiv preprint arXiv:2403.05527, 2024

    GEAR: An efficient KV cache compression recipe for near-lossless generative inference of LLMs. arXiv preprint arXiv:2403.05527. Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. 2023a.Efficient memory management for large language model serving with PagedAttention. arXiv preprint...

    work page arXiv 2025

  2. [2]

    decompress then dot-product

    Compressed-domain kernels. Replace “decompress then dot-product” with angle-domain com- pute fused into IO-aware decoding kernels, reducing HBM traffic per token under FlashAttention- style dataflow constraints (Dao et al., 2022; Dao, 2023; Ye et al., 2025)

    2022

  3. [3]

    decompress- then-dot-product

    Unified budgeting via rate–distortion control. Unify keep/drop decisions (Bui et al., 2025; Huang et al., 2025) with precision allocation(Xia et al., 2025; Su et al., 2025a) by allocating a fixed budget per token/head/layer over(i) retentionand (ii) angular/radial bits, optimizing an explicit distortion surrogate tied to downstream loss. Evaluation requir...

    2025

  4. [4]

    KV quantization as usual

    and paged KV serving ( Kwon et al., 2023b). No-reconstruction invariant (kernel-realized). The central systems constraint is an explicit invari- ant: No-reconstruction invariant. The decode kernel computes logits using [cos θ = f (cθ q, cθ k; b) and radius rk, without materializing any dense ˜k2 Rd in HBM. Accordingly, Spherical KV uses the angle-domain l...

    2022

  5. [5]

    Compute ∆i(b) (or a proxy) for each state and tier, along with cost(b)

    Score. Compute ∆i(b) (or a proxy) for each state and tier, along with cost(b). Enforce protection constraints immediately

  6. [6]

    Start from a safe tier for all kept states; then down-tier states with lowest vpB until the budget is met; only then drop the least valuable states

    Allocate tiers before dropping. Start from a safe tier for all kept states; then down-tier states with lowest vpB until the budget is met; only then drop the least valuable states

  7. [7]

    down-tier then drop

    Write tier-homogeneous pages. Group states by tier into contiguous pages with fixed headers, enabling specialized kernels and coalesced reads. This “down-tier then drop” structure is important: it avoids brittle discontinuities induced by hard truncation and typically yields smoother quality degradation under tighter budgets. B.3 Working Together: One Con...

    2022

  8. [8]

    it is measured under the same paged/ragged substrate and the same decode settings as the ref- erence,

  9. [9]

    it satisfies the same dense-relative quality tolerance∆,

  10. [10]

    This is why the frontier points are not merely visualization artifacts; they are the actual feasible operating states of the deployment system under an explicit fairness contract

    it carries a complete measured tuple (bKV, bHBM, s, q), with end-to-end accounting. This is why the frontier points are not merely visualization artifacts; they are the actual feasible operating states of the deployment system under an explicit fairness contract. D.1.2 Main frontier result across models and context lengths Structure of the existing 3 3 fr...

  11. [11]

    fixing the quality regime viaP ∆ m ,

  12. [12]

    fixing the deployment substrate via shared paged/ragged serving,

  13. [13]

    This is why Fig

    displaying the trade-off surface rather than isolated cherry-picked points. This is why Fig. 4 is the correct top-level result figure for the appendix. 500 750 1000 1250 1500 1750 2000 2250 200 250 300decode throughput (tok/s) Llama-3.1-8B-Instruct | L=8K 500 1000 1500 2000 2500 125 150 175 200 225 Qwen2.5-14B-Instruct | L=8K 500 1000 1500 2000 2500 150 1...

    2000

  14. [14]

    Γs is computed only overP ∆ m , so it does not reward hidden quality degrada- tion

    Quality conditioning. Γs is computed only overP ∆ m , so it does not reward hidden quality degrada- tion

  15. [15]

    Γm is computed only over points that meet or exceed dense throughput, so it does not reward trivially slow but memory-thrifty settings

    Speed conditioning. Γm is computed only over points that meet or exceed dense throughput, so it does not reward trivially slow but memory-thrifty settings

  16. [16]

    Both quantities use the same deployment-true definition of bKV, including paging and metadata overhead

    Shared accounting. Both quantities use the same deployment-true definition of bKV, including paging and metadata overhead

  17. [17]

    the stored rep- resentation is smaller

    Shared substrate. Both are measured under the same paged/ragged serving substrate and the same end-to-end runtime contract. For a systems paper, these properties are essential; without them, a reported gain can simply reflect a weaker operating condition rather than a genuinely better method. Why they matter more than raw single-point speedups. A raw sing...

    2022

  18. [18]

    transient runtime noise or cache luck,

  19. [19]

    a shifted cost hidden in another stage,

  20. [20]

    streaming + fused

    or an unfair comparison in kernel/runtime configuration. By contrast, when lower bKV, lower bHBM, and higher s co-occur under fixed decode settings and a shared substrate, the systems argument becomes much harder to challenge. Measured witness tuple. For every representative operating point and every frontier read-off, the appendix therefore treats the sy...

  21. [21]

    KV blocks are read once from HBM in a page-local pattern, with pointer-table resolution but without repeated revisits to the same block for staging or reconstruc- tion

    One-pass block-local reads. KV blocks are read once from HBM in a page-local pattern, with pointer-table resolution but without repeated revisits to the same block for staging or reconstruc- tion

  22. [22]

    The decode path does not write intermediate dense keys or auxiliary reconstructed representations to HBM

    No auxiliary staging buffers. The decode path does not write intermediate dense keys or auxiliary reconstructed representations to HBM

  23. [23]

    No dense key materialization.Similarity is computed directly from the stored representation, not from a reconstructed dense surrogate

  24. [24]

    streaming + fused

    Shared paged substrate. The same paged/ragged KV substrate is used for all compared methods, so differences in traffic arise from representation consumption rather than from different serving backends. Formalizing one-pass block-local reads. Let the KV cache be partitioned into pages of size P to- kens, indexed by a pointer table. For decode step t, letGt...

    2025

  25. [25]

    Similarity is computed directly from stored radial/angular codes rather than from reconstructed dense keys

    Angle-domain compute. Similarity is computed directly from stored radial/angular codes rather than from reconstructed dense keys

  26. [26]

    The controller jointly chooses retention and precision under a hard resident-memory budget

    Joint keep/drop + tiering. The controller jointly chooses retention and precision under a hard resident-memory budget

  27. [27]

    The emitted cache layout respects tier-homogeneous pages and block- local decode consumption

    Paged write contract. The emitted cache layout respects tier-homogeneous pages and block- local decode consumption

  28. [28]

    extra overhead

    No reconstruction. Dense key materialization is forbidden in the decode hot loop. Figure 11: Ablations A0–A5 at extreme context (L=128K): frontier shift + mechanism evidence under iso- quality (∆ = 0 .8). Top row (A0–A3, three LLMs): For each model (columns), we plot decode throughput s (tok/s; higher is better) versus effective resident KV budget bKV (by...

  29. [29]

    the gain is explained entirely by a better compute path

  30. [30]

    the gain is explained entirely by a better memory-control policy

  31. [31]

    What non-additivity means here

    the gain is non-additive, i.e., it arises because the compute path and the control policy reinforce one another under paged serving. What non-additivity means here. In this paper,non-additivity means that the best operating points reached by the joint system cannot be recovered by simply combining the best single-lever variants independently. Formally, a ...

  32. [32]

    A1 certification: direct compressed-domain compute reduces the densification tax relative to reconstruction

  33. [33]

    A2 certification: joint keep/drop + tiering outperforms retention-only, quant-only, and decoupled control at the same effective budgets

  34. [34]

    lost in the middle

    A3 certification: adaptive allocation structure is useful under the extreme-context regime rather than only at short lengths. Thus, the 128K ablation figure is not merely a difficult-case illustration. It is the appendix’s strongest causal evidence that the frontier shift survives precisely where the deployment problem is hardest. Table 6: Ablation matrix...

    2024

  35. [35]

    Increase ti so that the state is represented with more bits / less angular or radial distortion

    Tier escalation. Increase ti so that the state is represented with more bits / less angular or radial distortion

  36. [36]

    Override a drop decision zi = 0 and pin the state in cache when its local risk is too high

    Forced retention. Override a drop decision zi = 0 and pin the state in cache when its local risk is too high. The important point is that both interventions arepredictive. The gate is not explaining failures after they occur; it is altering the memory policy when a local drift estimate signals that a future failure is plausible. What A5 is meant to prove....

  37. [37]

    it ensures that the controller is responsive to genuinely dangerous local events

  38. [38]

    Thus the gate is not merely a safety heuristic

    it ensures that the controller is stable as a systems mechanism, avoiding thrashing and excessive control overhead. Thus the gate is not merely a safety heuristic. It is part of the systems design: it stabilizes the con- troller itself while stabilizing the model’s decode trajectory. How this ties back to the frontier and failure results. The stability fi...

  39. [39]

    choose the point with larger s(p),

  40. [40]

    if still tied, choose the point with smaller bKV(p),

  41. [41]

    if still tied, choose the point with smaller bHBM(p),

  42. [42]

    This rule ensures that every panel yields exactly one representative point and that the same point will be recovered whenever the sweep is rerun under the same protocol

    if still tied, choose the point with the smallest budget index B in the sweep order. This rule ensures that every panel yields exactly one representative point and that the same point will be recovered whenever the sweep is rerun under the same protocol. Interpretation of the ⋆ marker. The ⋆ marker in each frontier panel is therefore not decorative. It de...

  43. [43]

    the Dense-KV reference, represented by the best dense-quality point Q⋆ dense under the same de- code settings,

  44. [44]

    The table therefore provides a panel-by-panel numerical rendering of the frontier claim

    the representative Spherical-KV point p⋆, selected by the deterministic rule above from the re- tained frontier. The table therefore provides a panel-by-panel numerical rendering of the frontier claim. Quantities reported in the read-off table. Each row reports: • Q⋆ dense: best Dense-KV quality under the panel’s decode settings; • QSphKV: quality at the ...

  45. [45]

    quality-constrained byP ∆ m ,

  46. [46]

    frontier-constrained by Env∆ m,

  47. [47]

    objective-driven through s/bKV,

  48. [48]

    Thus, Table 7 is not a replacement for the frontier; it is the reproducible numerical anchor derived from it

    deterministic through an explicit tie-breaking rule. Thus, Table 7 is not a replacement for the frontier; it is the reproducible numerical anchor derived from it. D.5.4 What the representative-point table certifies What Table 7 certifies. Taken as a whole, the representative-point table certifies five things. (i) Matched quality. Every Spherical-KV row sa...

  49. [49]

    a minimal trigger: the smallest reproducible condition under which the failure appears,

  50. [50]

    a local mechanism: the model-side or controller-side reason the failure occurs,

  51. [51]

    a witness signal : an observable instability metric, such as a quality drop, Straj ", ∆T ", or disagreement/failure-rate increase,

  52. [52]

    The purpose of this section is therefore diagnostic and causal

    a controller intervention: a concrete fix, such as tier escalation, protected retention, or outlier- safe handling, together with its bounded systems cost. The purpose of this section is therefore diagnostic and causal. It shows that the safeguards in Spher- ical KV are predictive and actionable: they identify concrete failure regimes before they become w...

  53. [53]

    evidence allocation failures (retrieval distractor confusion),

  54. [54]

    decision-boundary failures (small-margin step flip, refusal flip),

  55. [55]

    outlier-amplified failures (outlier amplification),

  56. [56]

    termination and structure failures (termination drift, structured-output corruption),

  57. [57]

    just KV compression

    action-format failures (tool corruption). This coverage is sufficient for the appendix’s purpose because it shows that the controller addresses not just one type of degradation, but the main ways in which local approximation can become a deployment-relevant behavioral error. A longer list would add examples, but not much additional structure. Why the safe...

  58. [58]

    a kernel-native compressed-domain attention path(ADA),

  59. [59]

    a budget-explicit rate–distortion controller(RDR),

  60. [60]

    fast enough and accurate enough on average

    a stability release protocolbased on explicit witness tests. These three objects matter separately, and the results suggest that the paper’s practical value comes from their combination rather than from any one of them alone. (i) Kernel-native compressed-domain attention. The most important distinction between Spheri- cal KV and a large fraction of prior ...

  61. [61]

    kernel realism: does the runtime actually realize the claimed memory reduction as lower decode- time traffic, rather than silently paying a reconstruction tax?

  62. [62]

    A method that is fast but not stable is not usable; a method that is stable but not kernel-realized is not an actual serving win

    behavioral stability: does the operating point preserve length, stop behavior, seed stability, schema validity, tool correctness, and related behavioral invariants strongly enough for deploy- ment? The paper’s systems claim only holds when both constraints are satisfied simultaneously. A method that is fast but not stable is not usable; a method that is s...

  63. [63]

    KV traffic dominates wall-clock at the target context lengths, so further reductions in bytes/token are directly valuable

  64. [64]

    batch concurrency is limited by KV residency, making resident-memory reduction operationally useful even before peak context is reached

  65. [65]

    workloads are retrieval-heavy or tool-sensitive, so graceful degradation and witness-gated sta- bility matter as much as raw average accuracy

  66. [66]

    In these settings, the method is not merely plausible but particularly well-matched to the deploy- ment bottleneck the paper targets

    the serving stack already uses paging/offloading, so a further reduction in bytes/token can com- pose with existing memory-management strategies. In these settings, the method is not merely plausible but particularly well-matched to the deploy- ment bottleneck the paper targets. Where the method may be less compelling. Conversely, the method is less likel...

  67. [67]

    fix a budget target (bytes/token or total KV GiB),

  68. [68]

    sweep tier allocations and retention settings to trace the frontier,

  69. [69]

    filter candidate points through the witness suite,

  70. [70]

    always use Spherical KV

    report only the surviving points together with the declaredkernel path, engine settings, and mem- ory accounting. This protocol is intentionally conservative. It discourages best-case reporting and makes the choice of operating point reproducible across engines and workloads. What practitioners should take away. The practical lesson is therefore not “alwa...