Blast Radius

Hope Mogale; MY Pitsane

REVIEW 5 major objections 4 minor 31 references

Blast Radius

T0 review · 5 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper claims that reversible eviction of dead context cuts token consumption in agentic coding loops by 17–26% while remaining byte-exact recoverable.

desk verdict Reversible burial plus recurrence-class RDM is a genuinely useful idea, but the reported 17-26% and the bounded-downside theorem are not supported by the deployed system or the evaluation as described. read the letter →

arxiv 2608.07440 v1 pith:DF5LGM5R submitted 2026-08-07 cs.AI

MY Pitsane , Hope Mogale This is my paper · ORCID

classification cs.AI

keywords agenticcodingtokenefficiencyreversiblecontextevictionrecurringdeadmatterentropyblastradiusmemorymanagementLaplaceruleofsuccession

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that the ballooning token cost of agentic coding loops can be cut by reversibly evicting dead context rather than by losing information through truncation or summarization. It argues that reversible forgetting has an asymmetric payoff: the worst-case cost of burying a body that later turns out to be needed is one bounded exhumation, while the saving from keeping it buried grows with every turn, so carrying provably dead context is dominated. The empirical claim is that a deployed version reduced tokens per task by 17–26% across seven OpenAI models, matched or beat lossy baselines in overflow rate, and stayed byte-exact reversible, with zero of 450 buried bodies ever recalled. A reader should care because the mechanism, if the deadness judgment is trustworthy, offers a general way to make agentic loops cheaper without sacrificing recoverability.

What carries the argument

The load-bearing mechanism is the reversible sweep operator, NECROPHORESIS. A body is not deleted but 'buried': it is archived verbatim to an on-device store and replaced in the active context by a compact scent skeleton (a mission name, file list, and archival key at fixed cost $\sigma \ll \tau$). Exhumation is a single keyed O(1) read that restores the body byte-exact, at bounded token overhead $\kappa$. The sweep is chosen as a min-cost knapsack cover over the gate-licensed candidate set, with greedy efficiency $e(b) = (\tau(b)-\sigma)/(q(b)\kappa+\varepsilon)$; recurrence classes are estimated by Laplace's rule of succession, $\hat{q}_c = (e_c+1)/(k_c+2)$, so routine transcripts that die every turn are buried on sight. Theorem 7.1 is the engine: it converts this operator into an asymmetric bet with bounded downside.

What would settle it

Construct episodes where a later feature request silently depends on a file read many turns earlier, run the policy, and count exhumations; if the realized exhumation rate exceeds the threshold $(\tau(b)-\sigma)\bar{m}/\kappa$ from Corollary 7.2, the domination claim is violated in that setting.

Watch

Extended reading notes

Core claim

The paper's central claim is that in an agentic coding loop, reversible forgetting dominates carrying dead context. The argument is Theorem 7.1: if a body with token cost $\tau$ is buried and replaced by a $\sigma$-token skeleton, the sweep policy's ledger relative to carry-all is $(\tau-\sigma)m' - \kappa\cdot\mathbf{1}[m<\infty]$, where $m'$ is the number of turns buried and $\kappa$ the cost of one exhumation; the worst case is a constant $-\kappa$, while savings grow linearly with burial duration. Therefore, by Corollary 7.2, burial has non-negative expected value whenever the body's resurrection probability $q(b)$ satisfies $q(b) \le (\tau(b)-\sigma)\bar{m}/\kappa$, a threshold the system measures from the burial/exhumation ledger. The paper further claims the mechanism works in practice: across seven OpenAI models, the full Blast Radius plus Recurring Dead Matter policy cut median tokens per task from 43,053 to 34,518 (17–26% per model), matched the token economy of truncation while remaining byte-exact reversible, produced the lowest overflow rate, and recorded zero exhumations among 450 buried bodies, 378 of them recurring dead matter.

Load-bearing premise

The entire saving rests on trusting the gate's verdict that a buried body is truly dead: the licensed candidates must carry essentially no information about the correct next action, and the experiments scripted tasks in which old context was provably dead, so a body wrongly labeled dead and later needed would not have been detected.

Editorial extensions

If this is right

Agentic coding sessions can reclaim 17–26% of submitted tokens per model, matching the token economy of lossy truncation while retaining byte-exact reversibility.
A wrong eviction costs at most one bounded exhumation of $\kappa$ tokens, regardless of how long the body was buried, so misclassifying live context as dead is cheap rather than catastrophic.
Recurring routine transcripts—build logs, test reruns, version-control checks—can be buried on sight via the Laplace rule of succession, because their estimated resurrection probability falls as $1/(n+2)$ with recurrence count.
The burial/exhumation ledger gives a running estimate of the realized resurrection rate, so the policy can detect when its deadness assumption is wrong and shrink the candidate set.
The two-channel reach estimate (temporal and structural) is formulated over a Polish space of context objects, so the same operators apply to other agentic settings beyond coding.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Our reading: the zero-exhumation result is an artifact of the scripted task suite, where old context was provably dead by construction; in unscripted coding, the gate's precision—not the burial operator—will determine whether the 17–26% saving survives.
Our reading: the Laplace rule of succession converts a subjective death judgment into a measured, self-correcting probability, and the decisive test of the method is whether exhumation telemetry stays near zero when tasks are not designed to make context dead.
Our reading: the Polish-space formulation points to a natural next step the paper leaves implicit—replace the hard $q=0$ classifier with a learned estimator trained on exhumation telemetry, allowing graded deadness.
Our reading: if reversibility is as cheap as claimed, token-billing plans could offer reversible memory as a product, reclaiming tokens today with a bounded-fee resurrection guarantee; that is a direct economic corollary of Theorem 7.1.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The real contribution here is the combination of byte-exact reversible eviction at message granularity with recurrence-class burial driven by Laplace's rule of succession. That is a sensible way to attack repeatedly injected routine transcripts in agentic coding loops, and I haven't seen that exact pairing in the cited prior work. The paper is also honest in one important place: Section 8 clearly says the shipped policy is a reactive hard rule, and the framework's job is to justify that simple policy, not to claim the learnable estimator is running.

The central bounded-downside claim does not hold up as stated. Theorem 7.1's ledger arithmetic is correct, but it assumes an exhumation trigger: if a buried body is needed, the sweep policy performs exactly one exhumation before the turn runs. The deployed system has no described mechanism to notice a misclassified body and exhume it in time. A wrong burial isn't a bounded κ cost in that system; it's a failed turn or a missed constraint. The only thing shielding the bound is Assumption 7.4, imported from the self-cited HCRC report, and the evaluation never stresses it. The task suite is scripted so old context is provably dead, success is 100% everywhere, and zero of 450 buried bodies are exhumed. That zero is built in, not measured. RQ2's estimator accuracy is promised but never reported, and the "predictive" layer in the abstract is the reactive census of Section 8.

Other soft spots, in proportion: no code, data, or bootstrap intervals are actually released despite the preregistration language; the Polish-space formalism is decorative; and there's a small internal inconsistency—56 episode-runs in the protocol text versus 70 in the Table 2 caption, plus "both episodes" in Figure 5—that makes me wary of the exact numbers. None of these kill the engineering idea; they mean the paper's empirical claims and the bounded-downside guarantee are not currently supported.

Who gets value: people building context management for agent loops. The reversible-burial plus RDM framing is worth knowing about; the 17-26% number should not be quoted.

Recommendation: send it out, but require code, a real deployment with exhumation telemetry, and at least one condition where the deadness assumption can fail. If the authors show a misburial is detected and exhumed inside the bounded cost, the argument becomes credible. As is, it's a promising system sketch, not a demonstrated one.

Referee Report

5 major / 4 minor

Summary. The paper proposes Blast Radius, a memory-management layer for agentic coding loops, together with NECROPHORESIS, a reversible eviction operator that archives dead context verbatim and replaces it with a compact skeleton, and Recurring Dead Matter (RDM), which buries near-identical repeated transcripts using Laplace's rule of succession. A Polish-space formulation is used to present retention, recurrence, and eviction as measurable objects, and Theorem 7.1 argues that reversible forgetting has bounded downside: a mistaken burial costs at most one exhumation κ, while correct burial saves (τ−σ) per resident turn. The experimental section compares five context policies across seven OpenAI models and reports a 17–26% token reduction, the lowest overflow rate, and zero exhumations among 450 buried bodies. The paper concludes that reversible forgetting makes carrying dead context dominated and that Blast Radius is a safe scoping layer beneath the HCRC gate.

Significance. If the central claims held, the paper would make a useful contribution: reversible eviction with a bounded worst-case cost is a principled alternative to lossy truncation and summarization, and the authors have built a real implementation with a released harness and preregistered research questions. The accounting identity in Theorem 7.1 is sound conditional on its assumptions, and the archival design (byte-exact storage, ledger, transactional burial) is a sensible engineering contribution. However, the load-bearing guarantee is not established for the deployed system: the shipped census has no exhumation trigger, the ε-deadness assumption is imported from a self-cited technical report and is not validated by the evaluation, and the reported empirical numbers largely come from a harness-only variant that is not deployed. As it stands, the paper overclaims its central result.

major comments (5)

[Theorem 7.1 / §8] The bounded-downside guarantee is not realized by the deployed system because that system has no exhumation trigger. The proof of Theorem 7.1 assumes that if b is required at turn t+m, 'the sweep policy performs exactly one exhumation' before that turn runs, but Section 8 describes a reactive census with hard qt(b)=0 and no estimator that detects when a buried body becomes needed. If a body is mislabeled dead, the ledger never records an exhumation (Remark 7.3 counts only exhumations the system actually performs), and the realized cost is a failed turn, a missed constraint, or an overflow—not the bounded κ of the theorem. The abstract's 'remained byte-exact reversible' and the conclusion's 'bounded-downside bet' therefore apply to the theoretical operator, not to the shipped system.
[Assumption 7.4 / §9] The ε-deadness assumption on the HCRC-licensed candidate set is imported from the self-cited technical report [1] and is never independently validated, and the reported evaluation cannot validate it. Section 9 scripts the task suite so that early context is 'provably dead' by construction and every task is answerable from recent context; P2 explicitly concedes that success was 100% everywhere and the suite did not stress correctness. Under that design, zero exhumations in 450 burials (Table 3) is a property of the script, not evidence about the accuracy of the HCRC deadness judgment or about the mutual information I(Dt; A* | Ct\Dt) in Assumption 7.4. A test that could discriminate would need episodes in which a later turn actually requires a body that was buried; without such episodes, RQ3's 'calibration' conclusion is vacuous.
[Corollary 7.2 / §7] The domination threshold is vacuous in the regime the paper itself says 'typically' holds. Corollary 7.2 states that burial has non-negative expected value whenever q(b) ≤ (τ(b)-σ)m̄/κ, and then observes that the right-hand side 'typically exceeds 1,' in which case the inequality holds for every q(b)∈[0,1]. In that regime the threshold imposes no constraint, Proposition 4.8's licensing condition is satisfied by every candidate regardless of liveness, and the zero exhumation rate in Table 3 cannot be read as confirmation that the system is below a meaningful threshold. The paper should either report the actual value of (τ-σ)m̄/κ for the deployed geometry or stop describing the condition as a measured, falsifiable threshold.
[§9 (RQ2) / §8] The preregistered RQ2—estimator accuracy |B̂−B|—is not reported anywhere in Section 9, even though Definition 4.1 and the abstract describe Blast Radius as a predictive estimator. Section 8 states that the deployed eviction budget is 'reactive rather than predictive' and triggers on measured backlog Dt, so the paper provides no evidence for the predictive component at all. Either report the RQ2 results or remove the predictive claim from the abstract and introduction.
[Abstract / §8 'Status of RDM'] The abstract's headline numbers are attributed to the shipped Blast Radius, but they come from condition E, whose RDM mechanism is not deployed. Section 8 says integration of RDM into Chalk's shipped census and radar 'is underway,' and Table 2 shows the deployed census alone (D) achieves only a 39,568 vs 43,053 token median (8%) and has a higher overflow rate (5.14) than summarization (4.07). The 17-26% reduction and lowest-overflow claim are therefore results of the harness-only full policy, not of the system described in Section 8. The claims should be re-scoped or the system description updated.

minor comments (4)

[Theorem 7.1] The burial-duration count m′ appears to include the turn at which the body is required; if exhumation occurs before turn t+m, only m−1 turns of skeleton-vs-body savings accrue. The worst-case −κ bound is unchanged, but the displayed formula should be corrected.
[Table 3] The column dE[q]=0 conflates the realized exhumation frequency with the posterior mean; for a class with e_c=0, Proposition 6.6 gives q̂_c=1/(n_c+2)>0, so the text should distinguish the empirical rate from the estimator.
[§9 / Table 2] The episode count is inconsistent: the text says '56 episode-runs' while Table 2's caption says '70 episode-runs,' and Figure 5 says 'both episodes.' Please reconcile the number of runs and report bootstrap intervals as promised in the metrics paragraph, since Tables 2-3 contain no uncertainty measures.
[Definition 6.5] The ε-ball recurrence class of Definition 6.5 is not shown to coincide with or approximate the deployed normalization Σ-equivalence used in Definition 4.6; the paper asserts d(b,b′)=0 on the quotient but does not establish that the composite metric degenerates exactly on Σ-equivalent transcripts.

Circularity Check

1 steps flagged · score 4.0 of 10

Token-ledger arithmetic is self-contained, but the lossless-guarantee rests on Assumption 7.4, whose deadness judgment is imported from the same-authors' HCRC report [1] and whose zero-exhumation validation is scripted by the task design.

self citation load bearing [Sec. 7.1, Assumption 7.4 and following paragraph; cf. Sec. 2 'Blast Radius is deployed under HCRC [1]']
"Assumption 7.4 (ε-dead candidates). The HCRC-licensed candidate set D_t carries near-zero information about the correct next action: I(D_t; A*_{t+1} | C_t \ D_t) ≤ ε. ... Assumption 7.4 is exactly what the HCRC gate is for: it licenses a body into D_t only when verification has settled the sub-task the body records ... Blast Radius does not re-establish deadness; it consumes the gate's judgment."

Proposition 7.5, the formal bound that sweeping the HCRC-licensed set costs at most ε of task-relevant information, is I(Φ_S(C_t); A*) ≥ I(C_t; A*) − ε, which is true only under Assumption 7.4. The paper does not derive or independently validate ε-deadness; the only cited source is [1], a same-author technical report. Remark 7.3 claims the threshold is 'measured, not assumed' via exhumation telemetry, but Table 3's zero exhumations come from a task suite where early context is 'provably dead' and P2 admits every task is answerable from recent context, so the telemetry cannot test the assumption. The bounded-information-loss guarantee therefore reduces to the HCRC gate's self-cited judgment rather than to an independent derivation.

full rationale

The central accounting identity is not circular. Theorem 7.1 is a direct token-ledger computation: each buried turn saves τ(b)−σ and each required body costs one exhumation κ, so net saving is (τ(b)−σ)m′ − κ·1[m<∞] given the midden axioms and an actual exhumation before the needed turn. Corollary 7.2 is the expectation of that identity, and Proposition 4.8 is the standard Beta-Bernoulli posterior mean, not a fitted parameter. The 17–26% token reduction is read from measured submitted-token counts under fixed task conditions, so it is an external empirical result rather than a consequence of the equations. The circularity is localized to the information-loss-free guarantee. Proposition 7.5's bound holds only under Assumption 7.4, which asserts the HCRC-licensed candidate set is ε-dead; that assertion is imported from [1], a same-authors technical report, and the paper states it consumes the gate rather than re-establishing deadness. Remark 7.3's claim that the threshold is 'measured, not assumed' is weakened by the evaluation design: the task suite makes old context provably dead and P2 concedes success was 100% because each task is answerable from recent context, so zero exhumations is created by the script, not discovered. This is one load-bearing self-citation for the safety half of the claim, while the token-savings accounting retains independent content, so the score is 4 rather than higher.

Assumptions & free parameters 11 free parameters · 8 assumptions · 5 invented entities

The central claim depends on several hand-set constants (K, floor, σ, θ, radar parameters, tiers) and on unverified engineering axioms (byte-exact archival, bounded exhumation cost, ε-dead candidates). The formal Polish-space machinery is not operationalized: its metric weights and λ are unspecified, and the deployed system replaces it with a hard normalization map and a reactive threshold. Several new named constructs (NECROPHORESIS, RDM, midden, scent skeleton) are introduced with evidence only from the paper's own scripted harness.

free parameters (11)

K (protected window) = 3
Number of most recent missions presumed live; hand-chosen, not optimized. Sec 8.
candidacy floor = 800 chars
Minimum body length to be a burial candidate; design choice, Sec 8.
scent skeleton cost σ = ≈60 tokens
Fixed residual cost of a skeleton; assumed constant in the accounting theorems. Sec 8.
sweep threshold θ = 4000 tokens
Reclaimable-mass threshold triggering a recommended sweep; hand-chosen. Sec 8.
radar map constants (a, c, rmax) = (22, 3.1, 126)
Affine offset, square-root scale, and cap of the radar blip encoding; chosen for visual proportionality, not fitted. Sec 5.
churn tiers (θMED, θHIGH, θRISK, θDEADLY) = (50, 200, 500, 1000)
Risk bands for commit pressure; engineering thresholds. Sec 5.
safety margin γ = unspecified
Context-window safety margin in eviction budget, Def 4.2; no value given.
metric weights α, β, γ, δ = unspecified
Weights of the composite Polish metric, Def 6.1; required to sum to 1 but never set or used in deployment.
λ tuning constant = unspecified
Scales code-channel churn in blast-radius function, Def 6.2; never assigned, not used in the hard-rule deployment.
token ratio τ(b) ≈ |b|/4 = 0.25 token/char
Empirical token-per-character ratio used in the census; a modeling constant. Sec 8.
exhumation cost κ = small constant
Token overhead of one exhumation; assumed bounded but never measured or stated numerically. Assumption 4.4, Cor 7.2.

assumptions (8)

domain assumption Midden axioms: byte-exact archival (A2), auditability (A5), redaction (A7), bounded exhumation cost κ.
Assumption 4.4 states these as requirements the implementation is engineered to satisfy; the reversibility theorems depend on them.
domain assumption ε-dead candidates: HCRC-licensed candidate set Dt has near-zero mutual information with the correct next action.
Assumption 7.4; the bounded-information-loss proposition relies on this and it is never independently validated; HCRC [1] is a self-cited technical report.
standard math Correct next action and mutual information are well-defined for a session.
Section 7.1 assumes a correct next action A* and uses Shannon mutual information; this is a standard but idealized modeling assumption.
standard math Beta-Bernoulli posterior mean with uniform prior (Laplace rule of succession).
Prop 4.8 and Prop 6.6 use the standard posterior mean (e+1)/(n+2) for exchangeable Bernoulli trials.
ad hoc to paper The context space X is Polish with product decomposition and composite metric.
Definition 6.1 introduces a separable completely metrizable space of context objects; the metric weights are unspecified and the deployment replaces the metric with a hard normalization map Σ.
standard math Data-processing inequality applies to the skeleton transformation.
Used in Prop 7.5 to bound information loss; the skeleton is a deterministic function of the body.
domain assumption Dependency graph G and AST are available and correctly parsed.
Section 5 assumes the AST and dependency DAG are read from the repository; the code channel inherits this from the editor and executor.
domain assumption Token cost τ(b) is proportional to character count with ratio 1/4.
Sec 8 census estimates τ(b) ≈ |b|/4 chars; the ratio is an empirical approximation used to compute reclaimable mass.

invented entities (5)

NECROPHORESIS eviction operator
purpose: Reversible burial of dead context bodies into an archive with resident skeletons.
Named operator introduced by this paper; its behavior is defined internally and the only evidence is the paper's own harness.
Recurring Dead Matter (RDM) classification
purpose: Identifies near-identical repeated transcripts and buries all but the newest instance per class.
Introduced in Sec 4.4; the 378-of-450 count is from the paper's own scripted experiment, not independent data.
Midden (on-device archive)
purpose: Append-only byte-exact store for buried bodies plus a ledger.
A software component described in Sec 4.2 and Sec 8; the reversibility guarantee is an assumption (A2), not externally evidenced.
Scent skeleton
purpose: Compact resident stand-in naming a mission, touched files, and archival key.
A key-value representation introduced for the sweep operator; not independently validated.
Blast Radius estimator B
purpose: Predicts an incoming prompt's reach through context and code channels.
Defined in Def 4.1 and 6.2 but not deployed in the measured system; no RQ2 results validate it.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Blast Radius." pith.science (2026). https://pith.science/paper/DF5LGM5R

@misc{pith2026260807440,
  author       = {Pith},
  title        = {Pith review of: Blast Radius},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DF5LGM5R}},
  note         = {Machine review of arXiv:2608.07440}
}

read the original abstract

Agentic coding faces growing problems of affordability and wasted tokens. We introduce Blast Radius, a predictive memory management layer that estimates an incoming prompt's reach through coupled context and code channels. NECROPHORESIS enables reversible eviction by archiving dead context verbatim, while Recurring Dead Matter (RDM) identifies and buries repeatedly occurring transcripts. We formulate reversible context eviction over a Polish context space, providing a measurable foundation for retention, recurrence, and eviction while connecting context entropy to resurrection probability. Across seven OpenAI models, Blast Radius reduced token consumption by 17-26%, achieved the lowest overflow rate among tested policies, and remained byte exact reversible. Of 450 buried bodies, 378 were recurring dead matter and zero were recalled. Blast Radius operates beneath HCRC, determining which records to bury and how far an incoming prompt may reach into the codebase. This work contributes to the broader goal of Algosophy: making large language models and agentic coding more reusable and sustainable.

Figures

Figures reproduced from arXiv: 2608.07440 by the authors.

**Figure 1.** Blast Radius, two channels. An estimator Bˆ (1) scores the reach of prompt pt. The context channel (top) converts reach into an eviction budget Ht (2a); a knapsack sweep policy (3) buries the least-live bodies into the reversible midden M (4), reclaiming ∆ = P(τ − σ) tokens, each body exhumable in one O(1) read at bounded cost κ. The code channel (bottom) converts churn-weighted dependency reach Rk (2b) into a commi… view at source ↗

**Figure 2.** The burial lifecycle and its asymmetric ledger. Burial swaps the verbatim body (≈ τ tokens) for a scent skeleton (≈ σ ≪ τ ), reclaiming τ − σ on every resident turn; the body is archived byte-exact to the on-device midden (tables bodies and ledger), and exhumation restores it in one O(1) read at a one-time cost κ. Because M is lossless (A2), the operation is a bounded-downside bet, costing −κ in the worst case and s… view at source ↗

**Figure 3.** Operating pipeline. A turn’s edits are parsed to an AST (tree-sitter), which yields the churn weights w; the touched symbols enter the dependency DAG (the Coefficient DAG, whose nodes are Heaviside-gated); Blast Radius scores k-hop reach Rk over that DAG. The code channel emits the commit-pressure signal Πt when reach breaches the risk tier; the context channel emits the eviction budget Ht, which the supplementary N… view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: The Blast Radar: the code channel at full scale, wired into the AST and the DAG. Edits are parsed to the AST (upper left), whose per-file added/removed lines give the churn w(v) that sizes each blip; the touched files enter the dependency DAG (upper right, the Heavisid…

**Figure 5.** Figure 5: Context growth under the five policies. Mean tokens submitted at each turn, averaged over all seven models and both episodes. Carry-all grows without bound and lives above the context window (dashed) from mid-episode onward; truncation and summarization bend the curve …

**Figure 6.** Figure 6: Per-model token economy. Mean tokens submitted per episode for carry-all, the deployed Blast-Radius census, and the full policy with RDM, across the seven-model ladder. The RDM saving is uniform across model generations because routine traffic is a property of the sess…

**Figure 7.** Figure 7: Where the reclaimed mass comes from (condition E). Tokens reclaimed per episode, split into mission dead (the census of Sec. 4) and recurring dead matter (Sec. 4.4). Routine traffic, the same build log, test rerun, and version-control chatter dying every turn, contribu…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 20 canonical work pages

[1]

M. Y . Pitsane and H. Mogale. Heaviside Continuity of Rolling Coefficients for Eliminating Epistemic Entropy in Large Language Models. Technical report, Mankind Research Labs, 2026

work page 2026
[2]

G. Xiao, Y . Tian, B. Chen, S. Han, and M. Lewis. Efficient Streaming Language Models with Attention Sinks. InInternational Conference on Learning Representations (ICLR), 2024

work page 2024
[3]

Zhang, Y

Z. Zhang, Y . Sheng, T. Zhou, T. Chen, L. Zheng, R. Cai, Z. Song, Y . Tian, C. Ré, C. Barrett, Z. Wang, and B. Chen. H 2O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models. In Advances in Neural Information Processing Systems (NeurIPS), 2023

work page 2023
[4]

Z. Liu, A. Desai, F. Liao, W. Wang, V . Xie, Z. Xu, A. Kyrillidis, and A. Shrivastava. Scissorhands: Exploiting the Persistence of Importance Hypothesis for LLM KV Cache Compression at Test Time. In Advances in Neural Information Processing Systems (NeurIPS), 2023

work page 2023
[5]

S. Ge, Y . Zhang, L. Liu, M. Zhang, J. Han, and J. Gao. Model Tells You What to Discard: Adaptive KV Cache Compression for LLMs.arXiv preprint arXiv:2310.01801, 2023

arXiv 2023
[6]

T. Fu, H. Huang, X. Ning, G. Zhang, B. Chen, T. Wu, H. Wang, Z. Huang, S. Li, S. Yan, and G. Dai. Mixture of Attention Spans: Optimizing LLM Inference Efficiency with Heterogeneous Sliding-Window Lengths.arXiv preprint arXiv:2406.14909, 2024

arXiv 2024
[7]

Y . Ding, L. L. Zhang, C. Zhang, Y . Xu, N. Shang, J. Xu, F. Yang, and M. Yang. LongRoPE: Extending LLM Context Window Beyond 2 Million Tokens.arXiv preprint arXiv:2402.13753, 2024

arXiv 2024
[8]

Y . Wu, Y . Zhang, S. Ghosh, S. Basu, A. Deoras, J. Huan, and G. Gupta. ContextWeaver: Selective and Dependency-Structured Memory Construction for LLM Agents.arXiv preprint arXiv:2604.23069, 2026

arXiv 2026

Show all 31 references

[9]

Packer, S

C. Packer, S. Wooders, K. Lin, V . Fang, S. G. Patil, I. Stoica, and J. E. Gonzalez. MemGPT: Towards LLMs as Operating Systems.arXiv preprint arXiv:2310.08560, 2023

2023 arXiv
[10]

Chhikara, D

P. Chhikara, D. Khant, S. Aryan, T. Singh, and D. Yadav. Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory.arXiv preprint arXiv:2504.19413, 2025

2025 arXiv
[11]

S. Wang, E. Yu, O. Love, T. Zhang, T. Wong, S. Scargall, and C. Fan. MemMachine: A Ground-Truth- Preserving Memory System for Personalized AI Agents.arXiv preprint arXiv:2604.04853, 2026

2026 arXiv
[12]

Y . Wang, S. Cao, Y . Sun, H. Liu, Z. Hou, and X. Huang. Recursively Summarizing Enables Long-Term Dialogue Memory in Large Language Models.arXiv preprint arXiv:2308.15022, 2023. 17

2023 arXiv
[13]

Lewis, E

P. Lewis, E. Perez, A. Piktus, F. Petroni, V . Karpukhin, N. Goyal, H. Küttler, M. Lewis, W. Yih, T. Rock- täschel, S. Riedel, and D. Kiela. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. In Advances in Neural Information Processing Systems (NeurIPS), 2020

2020
[14]

Arslan, H

M. Arslan, H. Ghanem, S. Munawar, and C. Cruz. A Survey on RAG with LLMs.Procedia Computer Science, 246:3781–3790, 2024

2024
[15]

Cuconasu, G

F. Cuconasu, G. Trappolini, F. Siciliano, S. Filice, C. Campagnano, Y . Maarek, N. Tonellotto, and F. Silvestri. The Power of Noise: Redefining Retrieval for RAG Systems. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Re...

2024
[16]

Klesel and H

M. Klesel and H. F. Wittmann. Retrieval-Augmented Generation (RAG).Business & Information Systems Engineering, 67(4):551–561, 2025

2025
[17]

Y . Gao, Y . Xiong, Y . Zhong, Y . Bi, M. Xue, and H. Wang. Synergizing RAG and Reasoning: A Systematic Review.arXiv preprint arXiv:2504.15909, 2025

2025 arXiv
[18]

A. Asai, Z. Wu, Y . Wang, A. Sil, and H. Hajishirzi. Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection. InInternational Conference on Learning Representations (ICLR), pages 9112–9141, 2024

2024
[19]

R. S. Arnold and S. A. Bohner. Impact Analysis: Towards a Framework for Comparison. InProceedings of the International Conference on Software Maintenance, 1993

1993
[20]

Zimmermann, A

T. Zimmermann, A. Zeller, P. Weissgerber, and S. Diehl. Mining Version Histories to Guide Software Changes. InInternational Conference on Software Engineering (ICSE), 2004

2004
[21]

J. J. Flannery. The Relative Effectiveness of Some Common Graduated Point Symbols in the Presentation of Quantitative Data.Cartographica, 8(2):96–109, 1971

1971
[22]

Mogale and M

H. Mogale and M. Esiefarienrhe. Training and Optimizing Music Recommendation Algorithms Using Self-Similarity Matrices. EasyChair Preprint 7880, 2022

2022
[23]

M. Y . Pitsane, H. Mogale, and J. T. Janse van Rensburg. Improving Accuracy of Credit Card Fraud Detection Using Supervised Machine Learning Models and Dimension Reduction. InInternational Conference on Intelligent and Innovative Computing Applications (ICONIC), 2022

2022
[24]

H. Mogale. High Performance Canny Edge Detector using Parallel Patterns for Scalability on Modern Multicore Processors.arXiv preprint arXiv:1710.07745, 2017

2017 arXiv
[25]

C. E. Shannon. A Mathematical Theory of Communication.Bell System Technical Journal, 27(3):379–423, 1948

1948
[26]

J. Shi, Y . Zhu, X. Wang, Y . Kang, L. Wu, and T. Bai. Intrinsic Entropy of Context Length Scaling in Large Language Models.arXiv preprint arXiv:2502.01481, 2025

2025
[27]

Zhang, Y

Z. Zhang, Y . Wang, X. Huang, T. Fang, H. Zhang, C. Deng, S. Li, and D. Yu. Attention Entropy is a Key Factor: An Analysis of Parallel Context Encoding with Full-attention-based Pre-trained Language Models. InProceedings of the 63rd Annual Meeting of the ACL, pages 9840–9855, 2025

2025
[28]

Jiang, Q

H. Jiang, Q. Wu, X. Luo, C. Lin, Y . Yang, and L. Qiu. LLMLingua: Compressing Prompts for Accelerated Inference of Large Language Models. InProceedings of EMNLP, pages 13382–13403, 2023

2023
[29]

H. Li, Y . Zhang, N. Koto, Y . Qiang, Z. Yin, Y . Zhou, and I. W. Tsang. Compressing Context to Enhance Inference Efficiency of Large Language Models. InProceedings of EMNLP, pages 9715–9733, 2023

2023
[30]

L. Kuhn, S. Gal, and Y . Gal. Semantic Uncertainty: Linguistic Invariances for Uncertainty Estimation in Natural Language Generation. InInternational Conference on Learning Representations (ICLR), 2023

2023
[31]

Abbasi, M

R. Abbasi, M. G. Bellemare, R. Scandolo, and P. D. Velázquez. To Believe or Not to Believe Your LLM. arXiv preprint arXiv:2406.02543, 2024. 18

2024 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

M. Y . Pitsane and H. Mogale. Heaviside Continuity of Rolling Coefficients for Eliminating Epistemic Entropy in Large Language Models. Technical report, Mankind Research Labs, 2026

work page 2026

[2] [2]

G. Xiao, Y . Tian, B. Chen, S. Han, and M. Lewis. Efficient Streaming Language Models with Attention Sinks. InInternational Conference on Learning Representations (ICLR), 2024

work page 2024

[3] [3]

Zhang, Y

Z. Zhang, Y . Sheng, T. Zhou, T. Chen, L. Zheng, R. Cai, Z. Song, Y . Tian, C. Ré, C. Barrett, Z. Wang, and B. Chen. H 2O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models. In Advances in Neural Information Processing Systems (NeurIPS), 2023

work page 2023

[4] [4]

Z. Liu, A. Desai, F. Liao, W. Wang, V . Xie, Z. Xu, A. Kyrillidis, and A. Shrivastava. Scissorhands: Exploiting the Persistence of Importance Hypothesis for LLM KV Cache Compression at Test Time. In Advances in Neural Information Processing Systems (NeurIPS), 2023

work page 2023

[5] [5]

S. Ge, Y . Zhang, L. Liu, M. Zhang, J. Han, and J. Gao. Model Tells You What to Discard: Adaptive KV Cache Compression for LLMs.arXiv preprint arXiv:2310.01801, 2023

arXiv 2023

[6] [6]

T. Fu, H. Huang, X. Ning, G. Zhang, B. Chen, T. Wu, H. Wang, Z. Huang, S. Li, S. Yan, and G. Dai. Mixture of Attention Spans: Optimizing LLM Inference Efficiency with Heterogeneous Sliding-Window Lengths.arXiv preprint arXiv:2406.14909, 2024

arXiv 2024

[7] [7]

Y . Ding, L. L. Zhang, C. Zhang, Y . Xu, N. Shang, J. Xu, F. Yang, and M. Yang. LongRoPE: Extending LLM Context Window Beyond 2 Million Tokens.arXiv preprint arXiv:2402.13753, 2024

arXiv 2024

[8] [8]

Y . Wu, Y . Zhang, S. Ghosh, S. Basu, A. Deoras, J. Huan, and G. Gupta. ContextWeaver: Selective and Dependency-Structured Memory Construction for LLM Agents.arXiv preprint arXiv:2604.23069, 2026

arXiv 2026