Pith. sign in

REVIEW 5 major objections 9 minor 17 references

Proactive AI agents provably need fewer turns than reactive ones

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 · glm-5.2

2026-07-05 07:34 UTC pith:VWIJLMZU

load-bearing objection The Pareto improvement claim is true by construction — the theorems are correct but largely definitional, and load-bearing self-citations are unavailable. the 5 major comments →

arxiv 2605.23928 v1 pith:VWIJLMZU submitted 2026-04-21 cs.AI cs.CLcs.DCcs.MAcs.PLcs.SE

Context: Proactive Goal-Directed Intelligence via Composable Sandboxed Programs, Declarative Wiring, and Structured Interaction

Gregory Magarshak This is my paper
classification cs.AI cs.CLcs.DCcs.MAcs.PLcs.SE
keywords proactive dialoguegoal-directed agentswisdom librarysandboxed programsKV-cache optimizationorganizational efficiencyknowledge graphscross-platform governance
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.

This paper argues that conversational AI systems should be proactive—acting on their own when the next step is determinable from existing state—rather than waiting for user messages. It introduces an architecture called Context built on three mechanisms: write-time context assembly that precomputes typed attributes so context blocks are byte-identical across turns (enabling near-100% KV-cache reuse and roughly 10× cost reduction); a library of sandboxed imperative programs generated by language models, declaratively wired to goal types via typed stream relations and composed in a phase-ordered pipeline that preserves individual program correctness; and proactive goal stream state machines that inspect graph state and emit structured interaction content (option arrays, governance affordances, clarification prompts) without awaiting user input. The paper's central formal result is the Proactive Dominance Theorem: proactive agents weakly dominate reactive agents on expected turns to reach a terminal state, because any transition condition satisfiable by a background event (job completion, vote threshold, dependency resolution) can be acted on immediately rather than waiting for the next user message. Combined with a Quality Preservation argument—eliminated turns are coordination turns, not deliberation turns, so removing them does not degrade artifact quality—the paper claims a Pareto improvement: strictly faster, equal or higher quality, no tradeoff. Additional theorems establish correctness of program composition, soundness of declarative event wiring, cross-platform vote consistency, and a hierarchy-cache correspondence for multi-level artifact generation.

Core claim

The paper's load-bearing claim is that proactive agents—defined as agents that emit output whenever a graph-state condition is satisfied, not just when a user message arrives—provably require fewer expected interaction turns to complete a goal than reactive agents, and that the eliminated turns are coordination overhead (state inquiry, blocker identification, responsibility assignment, vote solicitation) rather than substantive deliberation. This makes proactivity a Pareto improvement: fewer turns, equal or higher quality, with no speed-quality tradeoff. The argument is structural rather than empirical: the turns removed are those where a reactive agent must wait for a user to trigger a step

What carries the argument

The central object is the proactive advancement condition: a triple (q, γ, μ) where q is a target state, γ is a predicate over stream attributes and graph state, and μ is a deterministic message generator. When the goal stream instance reaches state q and γ evaluates true on current graph state, the system emits structured content without any user input. This mechanism is what separates proactive from reactive agents and is what the dominance theorem turns on. Supporting it are the wisdom library (sandboxed imperative programs with phase ordering and typed I/O schemas), the policy graph for declarative event wiring, and the context block hierarchy for KV-cache reuse.

Load-bearing premise

The Quality Preservation Theorem assumes that artifact quality depends only on deliberation turns (where the language model makes a routing decision) and not on coordination turns. This means the turns eliminated by proactivity are defined as non-quality-contributing by construction. If coordination turns in real collaborations actually contribute to shared understanding, trust, or creative alignment, removing them could degrade quality and the Pareto improvement would not be

What would settle it

Measure artifact quality in multi-participant goal chats under proactive vs. reactive agents where coordination turns are independently rated for quality contribution. If coordination turns carry nonzero quality signal, the Pareto claim fails.

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

If this is right

  • If the dominance theorem holds in practice, multi-participant task chats (document review, support resolution, capability building) could see coordination overhead approach zero as the wisdom library matures, meaning fewer human turns spent on 'who does what' and 'what state are we in.'
  • The write-time context assembly approach—precomputing enriched attributes so context blocks are byte-identical between semantic changes—offers a structurally different cost model from RAG, where query-time retrieval produces variable context. Systems that can tolerate write-time latency could see substantial per-turn cost reductions.
  • The cross-platform vote consistency theorem means governance decisions (approvals, promotions, forks) can be shared across Telegram, email, web, and mobile with a single serialized ledger, enabling genuinely platform-agnostic organizational governance.
  • The program composition correctness theorem implies that a library of LM-generated programs can grow monotonically in coverage without breaking existing pipelines, provided addition-time phase-correctness validation is enforced—suggesting a path to self-improving agent systems.

Where Pith is reading between the lines

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

  • The Pareto improvement claim depends on a sharp distinction between coordination turns and deliberation turns. In real human-AI-human collaboration, some 'coordination' turns may build shared understanding, trust, or creative alignment that feeds back into artifact quality. If coordination turns have indirect quality contributions, the Pareto frontier may not be as clean as the formal model sugges
  • The dominance theorem is asymptotic in coverage: as the wisdom library and advancement conditions mature, savings grow. For early-stage systems with sparse coverage, the practical advantage over reactive agents may be modest, and the cost of building and maintaining the library may exceed the coordination savings it eliminates.
  • The architecture implicitly assumes that the correct next action is often determinable from graph state alone. For genuinely ambiguous or creative tasks, proactive actions based on state conditions could be premature or wrong, potentially introducing noise rather than reducing overhead. The paper acknowledges this with its domain-dependence discussion but does not bound the failure mode.
  • The independence assumption in the total efficiency gain formula (Corollary 6.5)—that coordination overhead elimination, LM-call elimination, and KV-cache savings act on independent cost components—is likely conservative in some directions and optimistic in others, since eliminated coordination turns also eliminate wisdom program invocations, creating coupling the product formula does not capture.

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

5 major / 9 minor

Summary. The paper presents Context, an architecture for proactive goal-directed AI agents that advance shared tasks without waiting for user prompts. The system rests on three mechanisms: (1) write-time context assembly with deterministic KV-cache reuse, (2) composable sandboxed 'wisdom programs' with phase-ordered composition, and (3) proactive state machine intelligence that emits structured content based on graph state. The paper proves six formal results: Context Stability (per-turn LM cost bound), Program Composition Correctness, Declarative Wiring Soundness, Proactive Dominance (proactive agents weakly dominate reactive agents on turns-to-terminal), Coordination Overhead Elimination and Quality Preservation (Pareto improvement), and Cross-Platform Vote Consistency. The architecture is implemented in the open-source Qbix/Safebox/Safebots stack.

Significance. The paper addresses a genuine gap: formal models of organizational efficiency in multi-participant human-AI-human goal-directed collaboration are rare, and the framing of proactive advancement as a structural rather than approximate property is novel. The wisdom library composition algebra (Theorem 4.2) and declarative wiring soundness (Theorem 4.5) are clean contributions. The open-source implementation across multiple platforms (Telegram, email, web, Apple) lends practical credibility. However, the central Pareto improvement claim (Theorems 5.6 + 6.3) is the headline result, and its significance depends on whether the quality preservation result is substantive or definitional, which is the primary concern discussed below.

major comments (5)
  1. §6, Theorem 6.3 (Quality Preservation): The theorem states 'Let Q(v) be any quality metric that depends only on the content produced by deliberation turns (turns with routing decision pass to the LM), not on coordination turns.' The proof then argues that since proactive advancement conditions fire on graph state (not LM generation), the eliminated turns are coordination turns, and since Q excludes coordination turns by definition, E[Q_πP] ≥ E[Q_πR] follows. This is correct but tautological: the theorem is true by construction because the quality metric is defined to exclude exactly the class of turns that proactivity eliminates. The Pareto claim ('strictly faster, equal or higher quality, no speed-quality tradeoff') is therefore definitional rather than substantive. This is load-bearing for the paper's central claim. The authors should either (a) acknowledge this explicitly and reframe
  2. the contribution as a definitional separation (coordination turns are structurally eliminable without touching deliberation turns) rather than a Pareto improvement, or (b) provide a theorem showing quality preservation under a broader class of quality metrics that could plausibly depend on coordination turns (e.g., metrics capturing shared understanding, trust, or participant buy-in). The current framing in §10 ('Domain dependence') acknowledges that c_elim is small for open-ended creative collaborations, but frames this as a coverage magnitude issue rather than as a threat to the quality preservation result itself.
  3. References [12] and [13] are cited as 'arXiv:2502.XXXXX' and 'arXiv:2501.XXXXX' respectively, with placeholder arXiv IDs. These references are load-bearing: the Byte-Identity Theorem [12], Accumulation Monotonicity [12], SPACER semantics [13], Local Linearizability [13], and RuleTrigger [13] are all invoked in proofs (Theorems 3.5, 4.5, 7.2, 8.1). Without verifiable prior work, the formal results that depend on these theorems cannot be fully assessed. If [12] and [13] are not yet publicly available, the authors should either (a) include the necessary definitions and theorem statements as an appendix, or (b) provide stable, accessible references. As it stands, a reader cannot verify the proofs of Theorems 3.5, 4.5, 7.2, or 8.1 without taking the cited results on faith.
  4. §6, Corollary 6.5 (Total Efficiency Gain): The product formula C_πP ≤ C_πR · (1−c_elim) · (1−E(W(t))) · [k_dyn + 0.1(k_perm + k_sess)] / [k_dyn + k_perm + k_sess] assumes independence of the three efficiency mechanisms. The authors acknowledge this is an approximation and that 'c_elim and E(W(t)) interact since eliminated coordination turns also eliminate wisdom library invocations.' However, the claim that the product 'provides a conservative upper bound under the approximation that the interactions are negligible' is not justified: if the interactions are non-negligible (as the authors themselves note), the product could either overestimate or underestimate the true savings depending on the direction of the interaction. The authors should either prove that the product is indeed conservative (i.e., an upper bound on savings) under stated assumptions, or remove the 'conservative upper'
  5. bound' claim and present the formula as a first-order approximation with explicit discussion of interaction effects.
minor comments (9)
  1. §3, Definition 3.3: The phase ordering relation lists 'pre ≺ ctx ≺ agg ≺ post ≺ render' and 'rel ≺ agg' and 'post ≺ auto' and 'idx independent.' The relationship between 'auto' and other phases (e.g., does auto ≺ render?) is not specified. Clarify the full partial order.
  2. §5, Theorem 5.6 proof: The phrase 'the expected number of turns between c_i becoming satisfiable and π_R acting on it is at least 1' should specify the expectation is over what distribution (presumably the arrival process of user messages). The current formulation is informal for a theorem proof.
  3. §5, Corollary 5.7: The variable p_user is described as 'the probability that a condition is satisfied simultaneously by a user message' but it is unclear whether this is per-condition or averaged across conditions. Clarify.
  4. §6, Theorem 6.2: The proof lists four categories C1–C4 but the theorem statement refers to 'categories C_1, ..., C_k' with general k. The proof should either argue that these four categories are exhaustive (and set k=4) or maintain generality in the proof.
  5. §7, Theorem 7.2: The proof references 'MM's Local Linearizability theorem [13]' and 'MM's per-publisher sequential consistency [13]' as distinct properties. It would help the reader to state these as formal assumptions (with self-contained definitions) rather than named references to unavailable work.
  6. §8, Theorem 8.1: The cost formula uses subscript ℓ for levels but the DAG H has 'd depth levels.' Clarify the relationship between ℓ and d, and whether N_ℓ is the number of leaf nodes at level ℓ.
  7. Table 1: 'Reqire' appears to be a typo for 'Require' in the SPACER phase description (also appears in §3, Definition 3.2: 'reads from a pre-loaded immutable input (no live DB queries, corresponding to SPACER's Compute; writes only via proposal accumulation (no direct writes, corresponding to Require)').
  8. §10 (Discussion): The proposed measurement methodology is described as 'the subject of ongoing work' but no concrete experimental design or timeline is given. Consider adding a brief description of planned experiments or at least specifying what goal types would be tested.
  9. The abstract lists six formal results but the introduction lists seven contributions (items 1–7). The seventh (Dual-traversal Hierarchy-Cache Correspondence, §8) is not mentioned in the abstract. Consider aligning.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for a careful and substantive review. The referee correctly identifies that Theorem 6.3 (Quality Preservation) is true by construction under the stated quality metric restriction, that references [12] and [13] contain placeholder arXiv IDs and need to be made verifiable, and that the 'conservative upper bound' language in Corollary 6.5 is not justified. We address each point below and describe concrete revisions.

read point-by-point responses

  1. Referee: Theorem 6.3 (Quality Preservation) is tautological: the quality metric is defined to exclude exactly the turns that proactivity eliminates, making the Pareto claim definitional rather than substantive. The referee requests either (a) explicit acknowledgment and reframing as a definitional separation, or (b) a theorem under broader quality metrics.

    Authors: The referee is correct that Theorem 6.3 is true by construction under the stated restriction on Q. We accept this criticism. In the revision, we will adopt option (a): we will explicitly acknowledge that the Quality Preservation result is a definitional separation—coordination turns are structurally eliminable without touching deliberation turns—rather than a substantive quality-preservation theorem under arbitrary quality metrics. We will reframe the contribution of Theorems 5.6 + 6.3 accordingly: the structural result is that proactive advancement conditions fire on graph state, not LM generation, so the eliminated turns are provably coordination turns, not deliberation turns. This is a non-trivial structural fact about the architecture, but it is not a universal quality-preservation guarantee. We will also add a paragraph in Section 6 discussing what a broader quality-preservation theorem would require (e.g., metrics capturing shared understanding or participant buy-in that could depend on coordination turns) and note that formalizing such a result would require additional assumptions about the relationship between coordination turns and terminal artifact quality, which we leave to future work. The informal argument for strict improvement (currently in the proof of Theorem 6.3) will be moved to a clearly labeled remark, separated from the formal theorem statement. revision: yes

  2. Referee: References [12] and [13] are cited with placeholder arXiv IDs and are load-bearing for proofs of Theorems 3.5, 4.5, 7.2, and 8.1. Without verifiable prior work, these proofs cannot be fully assessed. The referee requests either (a) including necessary definitions and theorem statements as an appendix, or (b) providing stable, accessible references.

    Authors: The referee is correct that placeholder arXiv IDs are unacceptable for load-bearing references. We will fix this by adopting option (a): we will add an appendix containing the necessary definitions and theorem statements from [12] and [13] that are invoked in our proofs. Specifically, the appendix will include: the Byte-Identity Theorem and Accumulation Monotonicity from [12] (needed for Theorems 3.5 and 8.1, and Corollary 4.3); the SPACER operational semantics, RuleTrigger delivery guarantee, and Local Linearizability theorem from [13] (needed for Theorems 4.5 and 7.2). Each will be stated with sufficient definitions for a reader to verify our proofs without accessing the companion papers. We will also update the references with correct arXiv IDs if the companion papers are publicly available by submission; if not, the appendix ensures self-contained verifiability regardless. revision: yes

  3. Referee: Corollary 6.5 (Total Efficiency Gain): The product formula assumes independence of the three efficiency mechanisms. The claim that the product provides a 'conservative upper bound' is not justified since interactions could cause the product to overestimate or underestimate savings. The referee requests either proving the bound is conservative or removing the 'conservative upper bound' claim and presenting the formula as a first-order approximation.

    Authors: The referee is correct. We cannot justify the 'conservative upper bound' claim as stated. The interaction between c_elim and E(W(t))—eliminated coordination turns also eliminate wisdom library invocations—means the product could overestimate savings (if the interactions are synergistic, the true savings exceed the product) or underestimate them (if double-counting occurs). In the revision, we will remove the phrase 'conservative upper bound' and reframe Corollary 6.5 as a first-order approximation under an explicit independence assumption. We will add a discussion of the interaction effects: (1) the c_elim × E(W(t)) interaction is synergistic (eliminated coordination turns eliminate their associated wisdom invocations, so true savings exceed the product), but (2) the c_elim × KV-cache interaction is partially redundant (eliminated turns already avoid their per-turn token cost, so the cache savings factor partially double-counts). The net direction of the combined interaction is not determined without empirical measurement. We will present the formula as an order-of-magnitude estimate and note that empirical validation is needed to assess the actual combined effect. revision: yes

Circularity Check

2 steps flagged

The Pareto improvement claim is true by construction: Theorem 6.3 defines quality to exclude exactly the turns that proactivity eliminates, then concludes quality is preserved.

specific steps
  1. self definitional [Theorem 6.3 (Quality Preservation Under Proactivity), §6]
    "Let Q(v) be any quality metric that depends only on the content produced by deliberation turns (turns with routing decision pass to the LM), not on coordination turns. Then E[Q_πP(v)] ≥ E[Q_πR(v)]."

    The theorem defines Q to depend ONLY on deliberation turns and explicitly NOT on coordination turns. Theorem 5.6 shows that proactivity eliminates coordination turns (turns where proactive advancement conditions fire on graph state, not LM generation). So the proof's chain is: (1) Q excludes coordination turns by definition, (2) proactivity only eliminates coordination turns, (3) therefore Q is preserved. This is tautological — the quality metric is constructed to be insensitive to exactly the dimension along which proactivity operates. The paper's central Pareto claim ('strictly faster, equal or higher quality, no speed-quality tradeoff') then follows from combining 5.6 (faster) with 6.3 (quality preserved), but 6.3 is true by construction. If coordination turns in real human-AI-human协作贡献

  2. self citation load bearing [References [12] and [13], cited throughout Theorems 3.5, 4.5, 7.2, 8.1, and Corollaries 4.3, 6.5]
    "[12] Gregory Magarshak. 2026. Grokers: Bottom-Up Inductive Comprehension and Write-Time Intelligence over Typed Knowledge Graphs. arXiv arXiv:2502.XXXXX (2026). [13] Gregory Magarshak. 2026. The Magarshak Machine: A Stream-Partitioned Model for Governed State Evolution. The SPACER Framework... arXiv arXiv:2501.XXXXX (2026)."

    Multiple load-bearing theorems depend on results from these two self-cited works: the Byte-Identity Theorem [12] underpins Theorem 3.5 (Context Stability) and Theorem 8.1; Accumulation Monotonicity [12] underpins Corollary 4.3 and Corollary 6.5; SPACER semantics [13] underpin Definition 3.2, Theorem 4.5 (Declarative Wiring Soundness), and Theorem 7.2 (Cross-Platform Vote Consistency); Local Linearizability [13] underpins Theorem 7.2. The arXiv IDs are placeholders ('2502.XXXXX', '2501.XXXXX'), suggesting these works are unpublished or unavailable for independent verification. The theorems that cite them are not self-contained — their correctness depends entirely on unverified claims from the same author's prior work.

full rationale

The paper's central organizational result — the Pareto improvement claim — rests on Theorem 6.3, which is true by construction: quality is defined to exclude coordination turns, and proactivity only eliminates coordination turns, so quality preservation follows tautologically. This is a genuine self-definitional circularity in the paper's headline claim. Separately, the paper's formal infrastructure relies heavily on two self-cited works [12, 13] with placeholder arXiv IDs, making several theorems (3.5, 4.5, 7.2, 8.1) dependent on unverified prior claims by the same author. However, not all theorems are circular: Theorem 5.6 (Proactive Dominance) is a straightforward consequence of its definitions (reactive agents wait for user messages, proactive agents don't), which is a legitimate formalization even if unsurprising. Theorems 4.2 and 4.5 are self-contained given the SPACER assumptions. The circularity is concentrated in the Pareto claim, which is the paper's most prominent organizational efficiency result.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 3 invented entities

The axiom ledger reveals heavy dependence on two unavailable self-citations [12, 13] for foundational results. The key free parameters (c_elim, p_user, fitness scores) lack empirical grounding. The ad hoc axiom that quality depends only on deliberation turns is constructed to make the Pareto theorem true.

free parameters (4)
  • c_elim (coordination elimination coverage)
    Defined as a weighted sum of coverage values for coordination categories C1-C4. The paper states 'for a mature system with complete coverage... c_elim -> 1' but provides no method for measuring or fitting these coverage values from real data.
  • p_user (probability condition satisfied by user message)
    Introduced in Corollary 5.7 as the probability that a transition condition is satisfied simultaneously by a user message. No empirical basis or estimation method provided.
  • T_c (semantic change interval)
    Used in Theorem 3.5 as the mean time between semantic changes. Treated as a given parameter from the system dynamics but no method for measuring it is provided.
  • Fitness score f_i for wisdom programs = [0,1]
    Each wisdom program has a fitness score in [0,1] used for evolutionary selection. The paper does not specify how fitness is computed or what data it is fit to.
axioms (5)
  • domain assumption Byte-Identity Theorem from [12]
    Theorem 3.5 (Context Stability) relies on this theorem to claim byte-identical context blocks across turns. The cited paper [12] has a placeholder arXiv ID and may be unavailable.
  • domain assumption SPACER operational semantics from [13]
    Theorems 4.5, 7.2 rely on SPACER's at-most-once delivery, RuleTrigger, RuleExecute, and Local Linearizability. Cited paper [13] has a placeholder arXiv ID.
  • ad hoc to paper Quality metric Q(v) depends only on deliberation turns, not coordination turns
    Stated in Theorem 6.3. This axiom is load-bearing for the Pareto improvement claim and is constructed to make the theorem true by definition.
  • ad hoc to paper Independence of efficiency mechanisms (coordination overhead, LM-call frequency, per-call token cost)
    Corollary 6.5 assumes the three efficiency mechanisms act on independent components. The paper acknowledges this is an approximation.
  • domain assumption Accumulation Monotonicity from [12]
    Invoked in Corollary 4.3 to claim the wisdom library grows in both coverage and correctness simultaneously. Depends on unavailable self-citation.
invented entities (3)
  • Wisdom library no independent evidence
    purpose: A governed library of LM-generated imperative programs that execute at interaction time without further LM calls
    No empirical evidence is provided that such a library can be effectively maintained, that fitness-based evolution produces useful programs, or that phase-correctness validation is tractable in practice.
  • Groker agents no independent evidence
    purpose: Write-time comprehension agents that precompute enriched typed attributes
    Referenced from [12] which is unavailable. No independent validation of their effectiveness is provided in this paper.
  • Proactive advancement conditions (Pi_G) no independent evidence
    purpose: Conditions that trigger proactive agent output based on graph state without user input
    The existence of such conditions is postulated. The paper does not demonstrate that they can be specified for real-world goal types with sufficient coverage to make c_elim substantial.

pith-pipeline@v1.1.0-glm · 14277 in / 2392 out tokens · 303491 ms · 2026-07-05T07:34:46.725952+00:00 · methodology

0 comments
read the original abstract

We present Context, the intelligence layer of the Magarshak Architecture, which replaces reactive query-response chatbots with proactive goal-directed agents that advance shared tasks without waiting for user prompts. The architecture rests on three mutually reinforcing mechanisms. Write-time context assembly precomputes enriched typed attributes via Groker agents, assembling interaction context as a deterministic pure function of graph state; context blocks are byte-identical across turns between semantic changes, enabling near-100% KV-cache reuse. Composable sandboxed wisdom programs form a governed library of LM-generated imperative programs declaratively wired to goal types via typed stream relations, composed via phase ordering, and executed at interaction time without further LM calls. Proactive goal stream state machines drive conversations toward terminal states by inspecting graph state and emitting structured interaction content (option arrays, governance affordances, clarification prompts) without awaiting user input. We prove six formal results: the Context Stability Theorem, bounding per-turn LM cost as a function of semantic change rate; a Program Composition Correctness Theorem; a Declarative Wiring Soundness Theorem; the Proactive Dominance Theorem, proving proactive agents weakly dominate reactive agents on expected turns-to-terminal-state; Coordination Overhead Elimination and Quality Preservation, establishing Pareto improvements in multi-participant goal chats; and a Cross-Platform Vote Consistency Theorem. Implemented in the open-source Qbix / Safebox / Safebots stack.

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

17 extracted references · 17 canonical work pages · 7 internal anchors

  1. [1]

    Anthropic. 2024. Prompt Caching.Anthropic Documentation(2024). https: //docs.anthropic.com/en/docs/build-with-claude/prompt-caching

    work page 2024

  2. [2]

    Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, et al. 2021. Pro- gram Synthesis with Large Language Models.arXivarXiv:2108.07732 (2021)

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  3. [3]

    Erik Brynjolfsson, Danielle Li, and Lindsey R. Raymond. 2023. Generative AI at Work.NBER Working Paper31161 (2023)

    work page 2023

  4. [4]

    Paweł Budzianowski, Tsung-Hsien Wen, Bo-Hsiang Tseng, et al. 2019. Hello, GPT-2! How Close Are We to Human-Level Dialogue Systems?. InACL Work- shop on NLP for Conversational AI

    work page 2019

  5. [5]

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, et al. 2021. Evaluating Large Language Models Trained on Code.arXivarXiv:2107.03374 (2021)

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  6. [6]

    Yang Deng, Wenqiang Lei, Wai Lam, and Tat-Seng Chua. 2023. A Survey on Proactive Dialogue Systems: Problems, Methods, and Prospects.arXiv arXiv:2305.02750 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  7. [7]

    Darren Edge, Ha Trinh, Newman Cheng, et al. 2024. From Local to Global: A Graph RAG Approach to Query-Focused Summarization.arXiv arXiv:2404.16130 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  8. [8]

    Matthew Henderson. 2015. Machine Learning for Dialog State Tracking: A Re- view. InProceedings of The First International Workshop on Machine Learning in Spoken Language Processing

    work page 2015

  9. [9]

    Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu Zheng, Yuheng Cheng, Ceyao Zhang, et al. 2023. MetaGPT: Meta Programming for Multi-Agent Col- laborative Framework.arXivarXiv:2308.00352 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  10. [10]

    Patrick Lewis, Ethan Petriv, Aleksandra Piktus, et al. 2020. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.Advances in Neural Information Processing Systems33 (2020), 9459–9474

    work page 2020

  11. [11]

    Guohao Li, Hasan Abed Al Kader Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. 2023. CAMEL: Communicative Agents for “Mind” Explo- ration of Large Scale Language Model Society.arXivarXiv:2303.17760 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  12. [12]

    Gregory Magarshak. 2026. Grokers: Bottom-Up Inductive Comprehen- sion and Write-Time Intelligence over Typed Knowledge Graphs.arXiv arXiv:2502.XXXXX (2026)

    work page 2026

  13. [13]

    Gregory Magarshak. 2026. The Magarshak Machine: A Stream-Partitioned Model for Governed State Evolution. The SPACER Framework: Streams, Policy, Actions, Capabilities, Execution, and Relations in Reactive Distributed Systems. arXivarXiv:2501.XXXXX (2026)

    work page 2026

  14. [14]

    Shakked Noy and Whitney Zhang. 2023. Experimental Evidence on the Pro- ductivity Effects of Generative Artificial Intelligence.Science381, 6654 (2023), 187–192

    work page 2023

  15. [15]

    Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Shaokun Zhang, Erkang Zhu, Beibin Li, Li Jiang, Xiaoyun Zhang, and Chi Wang. 2024. AutoGen: En- abling Next-Gen LLM Applications via Multi-Agent Conversation Framework. arXivarXiv:2308.08155 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  16. [16]

    Jun Xu, Haifeng Wang, Zhengyu Niu, Hua Wu, and Wanxiang Che. 2019. Con- versational Graph Grounded Policy Learning for Open-Domain Conversation Generation. InACL

    work page 2019

  17. [17]

    Williams

    Steve Young, Milica Gašić, Blaise Thomson, and Jason D. Williams. 2013. POMDP-based statistical spoken dialog systems: A review.Proc. IEEE101, 5 (2013), 1160–1179

    work page 2013