{"id":"1a9a2ee8-912e-4d5e-a2ee-5641facf4a3d","arxiv_id":"2506.03320","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Continual learning should pivot from weight-update-based methods to continual compositionality and orchestration of foundation models and agents.","lead":"This position paper argues that continual learning remains essential for foundation models, and that continual compositionality and orchestration (CCO) is the most promising of three directions. It offers a roadmap for moving from static monolithic models to ecosystems of continually adapted and composed models.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The motivating evidence for CCO conflates within-model inference scaling with compositional orchestration; the priority claim rests on a misclassified premise and an untested assumption that orchestration gains will outpace parameter scaling.","rationale":"The reader's weakest assumption is that orchestration-level gains will continue to outpace parameter and pre-training scaling. I agree that this assumption is load-bearing, but my concern is more specific: the evidence the paper cites for that assumption does not actually support it, because the cited advances are within-model scaling or inference techniques rather than compositional orchestration. This is an internal consistency issue, not merely a disagreement with external consensus. The paper's taxonomy and survey coverage are useful, and the authors are honest about open problems such as error propagation and the static nature of current orchestration frameworks. However, the central claim that CCO is 'the most promising and necessary direction' overreaches relative to the evidence presented, and the 'continual' aspect of CCO is asserted without any empirical or prototype demonstration. Because this is a position paper rather than a claim of new results, the appropriate response is to require tempering of the claim and engagement with the misclassification, which is exactly what a conditional acceptance should demand. I therefore keep the reader's CONDITIONAL verdict unchanged.","tokens_in":26001,"tokens_out":3926,"duration_ms":47646,"concrete_test":"Analytically re-derive the Section 1 argument by tabulating refs [12], [15], and [16] and classifying each advance as (i) parameter or pre-training scaling, (ii) within-model inference-time compute or context scaling, or (iii) cross-model composition and orchestration. If categories (i) and (ii) cover all three references, the premise that these are 'forms of orchestration rather than scale' is false, and the priority claim should be downgraded to a hypothesis. As a complementary empirical check, run a non-stationary benchmark in which a fixed pool of expert modules is used with (a) a static orchestrator and (b) a continually updated orchestrator that adapts routing and recombination as new tasks arrive; if (a) matches or beats (b) on average accuracy and adaptation latency, the 'continual' component of CCO adds no measurable value, further undermining the priority claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim (Section 1) is that CCO is 'the most promising and necessary direction for future continual learning research.' The decisive premise is that recent FM progress is 'orchestration rather than scale,' supported by refs [12], [15], and [16]. On inspection, however, these references are within-model inference techniques: [12] is test-time compute scaling inside a single multimodal model; [15] is post-training of reasoning LLMs; [16] is a longer context window for a single model. None involves composing multiple models or agents, and none is continual in the sense the paper uses for CCO. Calling them 'forms of orchestration' is doing the heavy lifting: from those trends the paper infers 'practical limits of an end-to-end foundation model' and the need for 'a modular, dynamically reconfigurable approach.' Even if modular orchestration later proves beneficial, the cited evidence does not establish that it, rather than inference-time compute or context scaling, is the source of gains. The 'continual' component is also asserted, not tested: Section 4.3.3 concedes current orchestration frameworks (LangGraph, AutoGen) are 'largely static and predefined,' and the paper presents no benchmark or prototype showing that dynamically re-routing or recombining modules over time outperforms a static composition. The load-bearing assumption—orchestration-level adaptation will outpace parameter scaling—is therefore both unproven and supported by evidence that on its face points to scaling of inference compute and context within a monolithic model.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper is a position/survey paper on continual learning in the era of foundation models. It reviews the history of continual learning, organizes current work into three directions—continual pre-training (CPT), continual fine-tuning (CFT), and continual compositionality and orchestration (CCO)—and argues that CCO is the most promising and necessary direction for future research. Section 2 provides background on classical continual learning methods and scenarios; Section 3 motivates continual learning for foundation models; Section 4 surveys each of the three directions, with CPT and CFT receiving substantial literature coverage and CCO framed around multi-agent orchestration, memory modules, and communication protocols; Section 5 and the conclusion recommend that academic research should prioritize CCO. The paper contains no experiments, benchmarks, or formal derivations beyond a restatement of the LoRA update.","tokens_in":26302,"tokens_out":6231,"duration_ms":70812,"significance":"The paper is valuable as a structured overview of an active debate and as a research agenda. Its strengths are the breadth of the CPT/CFT survey, the explicit enumeration of open problems (e.g., the stability gap, bias drift during CPT, PEFT-based continual learning, model merging), and a genuinely different architectural proposal that goes beyond the usual 'continual learning for LLMs' framing: a decentralized ecosystem of continually interacting models. The CCO discussion connects to relevant prior work on agents, test-time inference, memory modules (Voyager, GITM, AppAgent), and model merging. However, the central priority claim is not tested: no experiment, benchmark, or formal model separates CCO from CPT/CFT, and the key motivating premise—that recent FM progress is 'orchestration rather than scale'—is supported by references that, on inspection, concern within-model inference techniques. The paper is best read as a position piece proposing a research direction, not as a validated result.","major_comments":[{"comment":"The premise that 'recent advances in FMs have primarily emerged not from additional computational resources, but rather through enhanced reasoning abilities [15] and longer context windows [16], both of which are forms of orchestration rather than scale' is not supported by the cited evidence. Reference [12] is a test-time compute scaling study within a single multimodal model; reference [15] is a survey of post-training reasoning LLMs; reference [16] reports a longer context window for one model. None of these involves composing multiple models or agents, and test-time compute scaling is literally a form of inference-time scaling rather than orchestration. Because this premise is the main argument for the conclusion that CCO is 'the most promising and necessary direction,' the paper should either re-classify these results as within-model inference-time computation and argue separately that cross-model coordination will dominate, or soften the claim to a hypothesis with explicit scope conditions.","section":"Section 1"},{"comment":"The paper asserts that CCO 'inherently supports high-frequency adaptation' and that current frameworks such as LangGraph and AutoGen are 'largely static and predefined,' but it presents no evidence that dynamically re-routing or recombining modules over time improves performance over a static composition. A concrete falsifiable design is needed: compare, on a sequence of tasks with known distribution shifts, (i) a static orchestration graph with fixed routing, (ii) the same modules with a learned dynamic router that updates over time, and (iii) a single monolithic FM scaled to comparable total compute; report accuracy, forgetting, and adaptation latency. Until such a comparison exists, the priority claim remains an untested assumption.","section":"Section 4.3.3"},{"comment":"The load-bearing assumption that orchestration-level adaptation will outpace further parameter scaling is not examined. Section 5 states that real-world change is 'expressed at the level of orchestration' and that repeated offline training is impractical, citing [15]; however, reference [15] is about post-training large models with RL, which itself consumes substantial computational resources. The paper offers no evidence that high-frequency re-composition—new tools, revised chains of thought, reorganized agents—will continue to yield larger gains than improved weight-level continual learning or larger scale. To make the claim defensible, the authors should state a conditional thesis (e.g., 'if gains from coordination continue to grow as module pools and context lengths grow, then CCO is the highest-value direction') and identify data or benchmarks that would falsify it.","section":"Section 5"},{"comment":"The boundary between CFT and CCO is not crisply defined, which makes the priority claim hard to test. Model merging and adaptive routing methods (TIES-MERGING, DARE, MagMax, Adaptive LoRA Merging) are described under CFT in Section 4.2.3, yet they also compose models or modules dynamically; Section 4.3.3 then proposes 'dynamic selection and composition' as a CCO contribution. If module merging and routing are already CFT tools, then CCO is not a separate direction but an application of CFT, and the statement that CCO is 'not merely advantageous - it is indispensable' loses its contrast. Please state explicit inclusion criteria—for example, whether the modules being composed are frozen, whether the composition itself is learned, and what time scale counts as 'continual'—that separate CCO from CFT.","section":"Sections 4.2.3 and 4.3"}],"minor_comments":[{"comment":"There is a duplicated word: 'Fine-tuning is attractive because it it far cheaper' should read 'because it is far cheaper.'","section":"Section 4.2.1"},{"comment":"There is a typo: 'while preserving data privay' should read 'while preserving data privacy.'","section":"Section 4.2.3"},{"comment":"Reference [14] is cited together with [13] as evidence of multi-agent frameworks, but [14] is a chain-of-thought paper, not a multi-agent framework; please re-cite or rephrase the sentence.","section":"Section 1"},{"comment":"The cost estimate for training models 'comparable to GPT-4.5' is attributed to [47], which is Hoffmann et al.'s compute-optimal scaling paper rather than a GPT-4.5 cost estimate; the citation should be updated or the claim softened.","section":"Section 3"},{"comment":"The term 'orchestration' is used in two different senses: within-model inference-time computation (following [12], [15], [16]) and cross-agent composition (Section 4.3). This ambiguity should be resolved, for example by distinguishing 'intra-model orchestration' from 'inter-model orchestration'.","section":"Section 1"}],"recommendation":"major_revision","confidential_remarks":"This is a position paper rather than an empirical study. If the journal's scope emphasizes validated results, the fit is questionable; a perspective or opinion format may be more appropriate. The main technical weakness is that the central priority claim is stated categorically while resting on a misclassification of within-model inference scaling as orchestration. The paper's survey content is useful and generally accurate, but the authors should be pushed to either add a proof-of-concept comparison or explicitly reframe the contribution as an untested research agenda with falsifiable predictions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe one thing to know: this is a position paper, not a results paper. It proposes a three-part taxonomy—Continual Pre-Training, Continual Fine-Tuning, and Continual Compositionality & Orchestration—and argues the field should prioritize the third. The taxonomy is genuinely useful packaging, and the survey portions on CPT and CFT are careful, well-cited, and accurate enough to serve as an entry point to the area. The paper also deserves credit for being honest about current limits: Section 4.3.3 explicitly concedes that existing orchestration frameworks like LangGraph and AutoGen are “largely static and predefined,” and it does not pretend CCO is already realized.\n\nThe soft spot is the load-bearing premise. To argue CCO over CPT and CFT, the paper claims recent FM progress has come from “orchestration rather than scale,” citing [12] (test-time compute scaling within a single multimodal model), [15] (post-training of reasoning LLMs), and [16] (longer context windows). None of these is multi-model composition or dynamic reconfiguration. Calling them forms of orchestration is doing the work. The paper also assumes orchestration-level gains will outpace parameter scaling, but it offers no benchmark, prototype, or falsifiable design showing that dynamically re-routing or recombining modules over time outperforms static composition. The “continual” component is asserted rather than tested. So the central priority claim is an opinion, not a result.\n\nThat is not a reason to dismiss the paper. The descriptive material is solid, the framing is useful, and the authors openly identify open problems like error propagation and task decomposition. The main fix needed is reframing: CCO should be presented as an open hypothesis that the field should test, not as the established most promising direction. There are also minor typos and occasional overstatements in the concluding sections.\n\nI would send this to peer review as a position/survey paper, with a clear request to temper the headline claims and engage with the costs and limits of orchestration. A good referee would focus on the evidence gap in Sections 1 and 5. It is not a desk reject—just a paper that needs to say what it can and cannot support.","headline":"A useful and readable position paper whose headline claim—that continual compositionality and orchestration is the most promising direction—goes beyond what the paper's own evidence supports.","tokens_in":26858,"tokens_out":2297,"would_cite":false,"duration_ms":30843,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Continual learning's future lies in composing and orchestrating many models, not in scaling one monolithic model.","keywords":["continual learning","foundation models","continual pre-training","continual fine-tuning","continual compositionality","orchestration","multi-agent systems","catastrophic forgetting"],"falsifier":"Run a controlled benchmark that spends the same compute budget either on continued pre-training or fine-tuning of one model or on composing and orchestrating a team of smaller models; if the single model consistently achieves equal or larger accuracy gains across tasks, the paper's priority claim collapses.","tokens_in":25839,"feed_emoji":"🧩","tokens_out":4679,"duration_ms":47322,"temperature":0.7,"pith_summary":"This paper argues that continual learning remains essential even when foundation models have access to internet-scale knowledge, for three reasons: models must keep absorbing new data (continual pre-training, CPT), they must specialize and personalize cheaply (continual fine-tuning, CFT), and they must be dynamically composed with other models to solve higher-level tasks (continual compositionality and orchestration, CCO). The paper's central contention is that CCO is the most promising and necessary of the three, because recent AI progress has come less from added compute than from orchestration-like mechanisms such as reasoning, longer context windows, and multi-agent coordination. If correct, the field should pivot from scaling or weight-level continual learning toward modular, decentralized ecosystems of continually evolving and interacting agents, making continual learning more relevant than ever.","feed_headline":"Orchestration, not scale, is continual learning's future","feed_subtitle":"A review argues that dynamically composing specialized models will matter more than retraining one monolithic model.","key_machinery":"The central object is Continual Compositionality & Orchestration (CCO), defined as the dynamic integration of multiple AI agents over time to solve higher-level tasks. It carries the argument by supplying the high-frequency adaptation loop that CPT and CFT lack: instead of adjusting a single network's weights, an orchestrator selects and composes modules, including models, experts, tools, and memories, on the fly, while continual learning contributes rapid adaptation and memory consolidation to stabilize useful compositions. Supporting evidence includes test-time scaling laws that favor added inference compute over parameter growth, advances in reasoning and long context windows, and the diversity argument that a diverse set of agents can outperform a sample of the best individual performers.","core_discovery":"The central claim is that continual learning in the foundation-model era splits into three directions, and that the third direction, Continual Compositionality & Orchestration, will mark the rebirth of continual learning. CPT keeps a large model current against knowledge staleness and distribution shift; CFT enables task-specific specialization and personalization without full retraining; CCO treats an AI system as a living assembly of modules, prompt routers, domain experts, tools, and episodic memories that can be composed, recombined, and adapted at high frequency. The authors contend that CCO is indispensable because real-world change often appears at the level of orchestration, such as new tools, revised chains of thought, or reorganized agent teams, requiring updates in minutes or seconds and making repeated offline training cycles impractical. The future they foresee is not a single static model but an ecosystem of continually evolving and interacting models.","pith_inferences":["A natural test of the priority claim is to measure, on standard benchmarks, whether a fixed budget spent on orchestrating a team of smaller models reliably beats the same budget spent on continued pre-training or fine-tuning of one larger model.","The emphasis on CCO implies that evaluation protocols should move beyond task-incremental and domain-incremental benchmarks toward dynamic composition benchmarks where the set of available modules and the orchestration graph change over time.","If orchestration gains continue to outpace parameter scaling, weight-level continual learning could become a supporting technology that supplies the modules to be composed, rather than the main locus of continual learning research."],"forward_implications":["If correct, research emphasis in continual learning should shift from weight-level methods to orchestration and composition.","Academic labs can contribute to CCO without owning huge compute, since CCO is where the paper places primary emphasis for research.","Foundation-model systems become living assemblies whose components can be updated, replaced, or recombined as new tools, regulations, or tasks appear.","Continual learning becomes a foundational requirement rather than an optional enhancement, repositioning the field from niche research to the core of AI's next paradigm.","Specialized models distilled for particular domains will be composed rather than replaced by monolithic generalists."],"supporting_citations":[{"why":"Supplies the framing of task-shift and time-shift forgetting and the challenges that motivate continual learning for large language models.","marker":"[9]"},{"why":"Provides evidence that test-time scaling yields larger accuracy gains than parameter growth, supporting CCO over monolithic scale.","marker":"[12]"},{"why":"Serves as an example of multi-agent collaboration where several models negotiate, critique, or divide labor.","marker":"[13]"},{"why":"Supports the claim that recent advances emerged through enhanced reasoning abilities, a form of orchestration rather than scale.","marker":"[15]"},{"why":"Shows that longer context windows count as orchestration rather than scale and enable long-context composition.","marker":"[16]"},{"why":"Provides the diversity argument that a diverse subset of agents can outperform a sample of the best individual performers, motivating CCO.","marker":"[17]"},{"why":"Surveys foundation-language-model continual learning and frames the CPT and CFT directions with their adaptation benefits.","marker":"[45]"},{"why":"Supplies the notion of model compositionality and modular architectures as a path to dynamically composing task-specific modules.","marker":"[54]"}],"fun_headline_variants":["Orchestration, not scale, is the future of continual learning","Continual learning's rebirth: orchestration of models, not one giant model","Future AI is an ecosystem of evolving, composable models","Compositionality: the key to continual learning's next era","From monolithic to modular: continual learning's new direction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes that orchestration-level gains, such as test-time scaling, multi-agent coordination, and longer context windows, will keep outpacing gains from training bigger or better single models, so the field should prioritize CCO over the other two directions.","fun_headline_variants_meta":{"raw":{"variants":["Orchestration, not scale, is the future of continual learning","Continual learning's rebirth: orchestration of models, not one giant model","Future AI is an ecosystem of evolving, composable models","Compositionality: the key to continual learning's next era","From monolithic to modular: continual learning's new direction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00091,"raw_usage":{"total_tokens":3941,"prompt_tokens":1003,"completion_tokens":2938,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":619,"completion_tokens_details":{"reasoning_tokens":2851}},"tokens_in":619,"tokens_out":2938,"duration_ms":23594,"temperature":1.0,"reasoning_tokens":2851,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:05:55.986236+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a controlled benchmark that spends the same compute budget either on continued pre-training or fine-tuning of one model or on composing and orchestrating a team of smaller models; if the single model consistently achieves equal or larger accuracy gains across tasks, the paper's priority claim collapses.","supporting_citations":[],"review_version":1}