Multi-Source Morsels: Granular Source Integration

• Multi-Source Morsels are a computation pattern that decomposes heterogeneous sources into smaller, alignable, and auditable units for precise integration.
• They apply across domains such as neural morphology, model merging, graph databases, and multi-source QA, emphasizing selective granularity over uniform pooling.
• Empirical studies show that selective decomposition and recombination improve accuracy, robustness, and scalability in tasks ranging from language processing to object detection.

Searching arXiv for the cited papers to ground the article in the current record. “Multi-Source Morsels” denotes a recurring multi-source computation pattern in which information from several sources is not treated as a single undifferentiated block, but is decomposed into smaller transferable, alignable, or auditable units that can be selected, merged, or compared with finer granularity. In neural morphology, those units are multiple inflected forms from the same lemma; in model merging, they are rank-one singular components of task updates; in graph database systems, they are dispatchable source or frontier morsels, including packed multi-source morsels; and in multi-source QA, the operative unit can be the relationship between source-grounded answers rather than a single gold answer (Kann et al., 2016, Osial et al., 26 Aug 2025, Chakraborty et al., 26 Aug 2025, Li et al., 27 May 2026, Yang et al., 28 May 2026). This suggests a common technical theme: multi-source performance depends less on mere source count than on how the system decomposes, prioritizes, and recombines source-specific evidence.

1. Conceptual structure of multi-source morsels

A recurrent difficulty in multi-source settings is that sources are heterogeneous, partially redundant, and sometimes conflicting. Several papers therefore reject uniform pooling. AXIS argues that prior multi-source transfer methods such as aTLAS are too coarse because they operate on full task vectors, layers, or partitions, which can mix irrelevant or noisy information and scale linearly with the number of sources (Osial et al., 26 Aug 2025). CMSS makes the analogous point for multi-source unsupervised domain adaptation: if all sources are forced to align equally, the process can induce negative transfer, especially when some latent source domains are far from the target (Yang et al., 2020). DMSN reports the same phenomenon for object detection, explicitly observing that directly combining sources can perform worse than using the best single source (Yao et al., 2021).

The compensating design move is selective granularity. In AXIS, each fine-tuned source update is decomposed into rank-one terms and only the top-$K$ singular components across all sources are retained (Osial et al., 26 Aug 2025). In CMSS, a curriculum manager reweights individual source samples over time, emphasizing those that are harder for the discriminator to separate from target and therefore more transferable (Yang et al., 2020). In DMSN, low-level features are strongly aligned, high-level features are weakly aligned, and source-specific subnets are preserved while a pseudo target subnet is learned by weighted parameter combination (Yao et al., 2021). In Kuzu, recursive query parallelism is re-expressed as a dispatch-policy design space over source morsels, frontier morsels, hybrid morsels, and multi-source morsels (Chakraborty et al., 26 Aug 2025).

A second recurring feature is that evaluation often shifts from endpoint accuracy alone to the quality of source interaction. In medical multi-source RAG, the unit of analysis becomes the inter-source relationship between answers grounded in different institutional handbooks (Li et al., 27 May 2026). In conflicting personal memory QA, the benchmark is explicitly constructed to separate source reachability, extraction quality, resolver quality, and abstention behavior (Yang et al., 28 May 2026). A plausible implication is that “morsels” are not only computational fragments; they are also epistemic fragments whose provenance and compatibility must remain visible.

2. Morphological reinflection as a canonical multi-source formulation

In neural morphology, multi-source morphological reinflection generalizes the standard single-source task by replacing one source form-tag pair with $k$ such pairs from the same lemma (Kann et al., 2016). The task is formalized as

$$\left(t_{\text{trg}}, \{(f^{1}_{\text{src}}, t^{1}_{\text{src}}), \ldots, (f^{k}_{\text{src}}, t^{k}_{\text{src}})\}\right) \mapsto f_{\text{trg}}.$$

Single-source MRI is the special case $k=1$. The motivation is that different cells in a paradigm may expose different stems, vowel alternations, or affixes, so no single observed form is necessarily sufficient.

The paper organizes this intuition into four source-utility regimes: AnyForm, SingleForm, MultiForm, and NoForm (Kann et al., 2016). The main empirical claim is that multi-source input is especially helpful in SingleForm cases, where one specific source form is predictive, and MultiForm cases, where multiple source forms are jointly necessary. This is tied directly to the linguistic notion of principle parts.

The architecture extends the earlier MED encoder-decoder model from one encoder to $k$ encoders with one shared decoder. Each source input contains source tag, source form, and target tag, represented at the character/subtag level. The decoder conditions on a context vector that attends jointly over all encoder states: $$c_t = \sum_{m=1}^{k}\sum_{i=1}^{|X_m|} \alpha_{mi}(s_{t-1}) h_{mi}, \qquad \sum_{m=1}^{k}\sum_{i=1}^{|X_m|} \alpha_{mi} = 1.$$ The main version uses shared encoder parameters across the $k$ encoders, and training minimizes the negative conditional log-likelihood of the gold target sequence (Kann et al., 2016).

The experiments extend the SIGMORPHON 2016 Shared Task data using UniMorph paradigms for Arabic, Finnish, Georgian, German, Hungarian, Russian, Spanish, and Turkish. For each original source-target example, three additional forms are uniformly sampled from the same paradigm, yielding a 5-tuple with 4 source forms and 1 target form. Evaluation uses 1-best exact-match accuracy (Kann et al., 2016).

The central result is that multi-source MRI outperforms single-source MRI for 7 of 8 languages. Using 2 sources is better than using 1 source in most languages, and using 4 sources is better than using 2 sources in all languages. The largest gains from one source to four sources are Arabic at +7.3% and Russian at +3.5%; Spanish gains +0.9% and Georgian +1.3%; Hungarian is the only exception, decreasing by about 0.3%, which the authors attribute to overfitting (Kann et al., 2016). Learning-curve analysis further shows that multi-source models generalize faster, need fewer training paradigms to reach a given accuracy, and help most in low-resource regimes; for Arabic, the difference can exceed 20% when only $1/8$ of the training data is used.

The morphological case provides a clean prototype of multi-source morsels: the relevant evidence is not “more data” in the aggregate, but a small set of complementary paradigm cells whose utility depends on the target form.

3. Rank-one knowledge morsels in model merging

AXIS reformulates multi-source transfer learning as model merging at the level of singular components rather than whole task vectors (Osial et al., 26 Aug 2025). The setting assumes a pretrained base model and many source models fine-tuned on different tasks; the goal is to adapt to a new target task without access to source data. The paper’s claim is that existing methods such as aTLAS remain coarse-grained, memory-heavy, and linearly dependent on the number of sources because they learn coefficients over entire task vectors, layers, or partitions.

For each matrix-valued source update $\Delta_i$, AXIS computes

$\Delta_i = U_i \Sigma_i V_i^\top,$

so that each source becomes a set of rank-one terms $k$0. The first stage aggregates all such terms from all source models into a global pool, ranks them jointly by singular value magnitude, and retains only the top-$k$1 most salient components: $k$2 Non-matrix parameters such as biases and normalization terms are averaged across sources, while the SVD-based procedure is applied independently per matrix layer (Osial et al., 26 Aug 2025).

Because the retained singular vectors come from different sources and are not orthogonal, AXIS then performs a second SVD,

$k$3

to obtain an orthogonalized transfer basis. Target adaptation does not fine-tune the whole merged update; it fine-tunes only the top-$k$4 principal singular values in $k$5, leaving all singular vectors and the remaining singular values frozen. The target objective is the usual cross-entropy loss, but optimization is restricted to these principal values: $k$6 The paper interprets large singular values as dominant structural changes that are more likely to encode transferable knowledge than small, noisy details (Osial et al., 26 Aug 2025).

Empirically, AXIS is evaluated on 21 vision target tasks using publicly released task vectors from CLIP ViT-B-32 and ViT-L-14. The tasks include DTD, EuroSAT, Flowers102, CIFARs, ImageNet, Food101, SUN397, PascalVOC, CUB200, GTSRB, and UCF101. Across all source-count regimes and parameter budgets, AXIS outperforms aTLAS in average Top-1 accuracy. With budgets corresponding to roughly 10%, 20%, and 40% of singular values, it consistently achieves higher accuracy and lower variance than aTLAS. For the ViT-B-32 setup, the trainable-parameter budgets are only a few thousand parameters; the same qualitative advantage is reported for ViT-L-14 (Osial et al., 26 Aug 2025).

The robustness findings are equally central. AXIS degrades much less than aTLAS when one source task vector is corrupted with large Gaussian noise, continues to transfer effectively when all source vectors are heavily pruned, and degrades more slowly under patch dropout and standard image corruptions, especially under moderate perturbations. Ablations show that the Stage-1 merged matrix $k$7, after a data-free rescaling of singular values, can already outperform naive averaging; that varying $k$8 leaves performance fairly stable; that selecting top singular components is better than selecting bottom or arbitrary components; and that skipping the final SVD orthogonalization causes a substantial drop, especially when many source models are aggregated (Osial et al., 26 Aug 2025).

Here the “morsel” is literal in the paper’s description: a fine-tuned weight update is recast as a collection of rank-one morsels of knowledge. The method’s contribution is to make those morsels globally comparable and cheaply adaptable.

4. Morsel dispatching in graph database systems

In recursive query evaluation for graph database management systems, “morsel” refers to a dispatchable unit of work rather than a piece of model knowledge (Chakraborty et al., 26 Aug 2025). The paper starts from an IFE-style serial execution model with two nested scans: an outer loop over source nodes and an inner loop over frontier nodes. It then defines a design space of morsel dispatching policies based on which loop is partitioned.

The source-node policy, 1T1S, assigns one whole source computation to one thread. It is simple and has no synchronization inside a source’s IFE execution, but it underutilizes hardware when the query has few sources. The frontier-level policy, nT1S, parallelizes the inner frontier expansion for a single source, which helps when one source has dense frontiers but is limited by sparse frontiers and by the fact that only some frontier levels may expose substantial parallelism (Chakraborty et al., 26 Aug 2025).

The hybrid policy, nTkS, combines both levels: up to $k$9 source morsels are active concurrently, and threads pull frontier morsels from whichever source morsel currently has work. The paper presents this as the robust default because it captures the good cases of both 1T1S and nT1S while outperforming them when either source count or frontier width becomes limiting (Chakraborty et al., 26 Aug 2025).

A further extension, nTkMS, applies the prior multi-source BFS optimization as a morsel-dispatch policy. Up to 64 sources are packed together in one multi-source morsel; each node stores a 64-bit lane mask, and if a node is active in several traversals then its adjacency list is scanned once and edgeCompute is applied to all active lanes. This reduces repeated adjacency scans, repeated neighbor activations, and memory traffic for neighbor exploration (Chakraborty et al., 26 Aug 2025).

The gains are workload-dependent. When the source count is high enough to saturate the 64 lanes, nTkMS improves performance by about 1.4x–4.4x. With fewer than 32 sources, however, nTkMS is often slower than nTkS, because the multi-source machinery adds overhead without enough scan sharing. Memory costs are explicit: for multi-source shortest paths with 64 lanes, frontiers and visited structures use 64 bits per node, the frontier/visited overhead is reported as 24 bytes per node per multi-source morsel, total upfront memory is 88 bytes per node for path lengths only, and 536 bytes per node for actual paths, plus extra thread-local memory for stored path edges. On Graph500-28, using 128+ sources while outputting paths can run out of memory (Chakraborty et al., 26 Aug 2025).

The graph-systems interpretation broadens the meaning of multi-source morsels. The essential issue is not transfer learning but dispatch granularity: the source library is transformed into work units whose size determines utilization, synchronization cost, scan reuse, and memory pressure.

5. Source-conditioned answers and conflicting evidence

In multi-source RAG and personal memory QA, the central problem is not how to merge parameters or dispatch parallel work, but how to audit or resolve source disagreement. In transplant patient education, the same question can yield different answers depending on which institutional handbook is retrieved. The paper argues that source-dependence is a missing axis of NLP evaluation and that the unit of evaluation should shift from answer correctness against one reference to the inter-source relationship between answers grounded in source $$\left(t_{\text{trg}}, \{(f^{1}_{\text{src}}, t^{1}_{\text{src}}), \ldots, (f^{k}_{\text{src}}, t^{k}_{\text{src}})\}\right) \mapsto f_{\text{trg}}.$$0 and source $$\left(t_{\text{trg}}, \{(f^{1}_{\text{src}}, t^{1}_{\text{src}}), \ldots, (f^{k}_{\text{src}}, t^{k}_{\text{src}})\}\right) \mapsto f_{\text{trg}}.$$1 (Li et al., 27 May 2026).

That argument is operationalized in TransplantQA, which pairs 1,115 real patient questions with 102 patient-education handbooks from 23 U.S. transplant centers spanning heart, kidney, liver, lung, and pancreas. The questions are split into 311 general questions, answerable by every handbook, and 804 organ-specific questions. Retrieval is handled by HERO-QA, a recall-first hierarchical retrieval system with structured extraction, length routing, four parallel first-stage retrievers, Reciprocal Rank Fusion, cross-encoder reranking, parent-section expansion, and low-evidence fallback. Answers are generated by Qwen3-32B with temperature 0 and a required NOT ADDRESSED behavior when relevant information is absent (Li et al., 27 May 2026).

The paper’s judge uses a validated 5-label taxonomy: Absent, Consistent, Complementary, Divergent, and Contradictory. Pairwise comparisons are stored in a matrix $$\left(t_{\text{trg}}, \{(f^{1}_{\text{src}}, t^{1}_{\text{src}}), \ldots, (f^{k}_{\text{src}}, t^{k}_{\text{src}})\}\right) \mapsto f_{\text{trg}}.$$2, and for the general subset each question yields $$\left(t_{\text{trg}}, \{(f^{1}_{\text{src}}, t^{1}_{\text{src}}), \ldots, (f^{k}_{\text{src}}, t^{k}_{\text{src}})\}\right) \mapsto f_{\text{trg}}.$$3 comparisons. Out of 5,730,465 pairwise comparisons, 4,519,245 (78.9%) are Absent; among non-absent pairs, Complementary accounts for 75.4%, Divergent for 12.9%, Consistent for 7.1%, and Contradictory for less than 0.1%. Better retrieval reduces absence by 13.6 percentage points on average, leaves per-pair divergence intensity roughly unchanged, and increases the fraction of questions with at least one divergent pair by 15.9 points. The paper’s interpretation is that earlier systems understated how often disagreement existed, not how intense it was (Li et al., 27 May 2026).

A parallel but more diagnostically controlled formulation appears in selective QA over conflicting multi-source personal memory (Yang et al., 28 May 2026). Here each persona has five evidence streams—profile_ltm, planner, daily_self_report, objective_log, and device_log—with controlled biases such as staleness, optimism, topic-dependent self-report bias, incompleteness, and missing sensor data. The benchmark contains 18 question templates across 8 reasoning types, 480 personas, 4 random seeds, and 34,560 instances. The model must answer from a discrete answer space or output SKIP, where SKIP is explicitly not part of the answer space (Yang et al., 28 May 2026).

The main answer-only macro accuracies show that a trained structured resolver outperforms prompt-only LLM baselines: DSNBF = 80.3%, NBF = 79.8%, SSB = 77.3%, while the strongest prompt-only baselines are around 70%, including Gemini 3.1 Pro Schema = 70.0% and GPT-5.4 Schema = 69.7%. In the selective setting, DSNBF reaches 85.3% selective accuracy at 78.3% coverage, whereas GPT-5.4 Schema reaches 71.0% selective accuracy at 95.4% coverage; the paper’s interpretation is that fusion methods expose calibrated posterior margins and can skip hard cases, while prompt-only LLMs answer almost everything and provide little usable calibration (Yang et al., 28 May 2026).

These works replace the notion of “best source” with a more stringent question: how stable is the answer under source substitution, and when should the system abstain because the morsels of evidence are incomplete or conflict-laden?

6. Selective source usage in adaptation and transfer

Multi-source morsel reasoning also appears in domain adaptation, where the difficulty is to decide which sources, or which parts of them, should influence target adaptation. CMSS addresses multi-source unsupervised domain adaptation when source samples come from multiple unknown domains and domain labels are unavailable (Yang et al., 2020). It adds an independent Curriculum Manager $$\left(t_{\text{trg}}, \{(f^{1}_{\text{src}}, t^{1}_{\text{src}}), \ldots, (f^{k}_{\text{src}}, t^{k}_{\text{src}})\}\right) \mapsto f_{\text{trg}}.$$4 to the usual feature extractor, classifier, and domain discriminator, and converts minibatch source samples into a softmax-normalized per-sample weighting. The weighted domain loss is

$$\left(t_{\text{trg}}, \{(f^{1}_{\text{src}}, t^{1}_{\text{src}}), \ldots, (f^{k}_{\text{src}}, t^{k}_{\text{src}})\}\right) \mapsto f_{\text{trg}}.$$5

with $$\left(t_{\text{trg}}, \{(f^{1}_{\text{src}}, t^{1}_{\text{src}}), \ldots, (f^{k}_{\text{src}}, t^{k}_{\text{src}})\}\right) \mapsto f_{\text{trg}}.$$6, and the curriculum is updated dynamically during adversarial training (Yang et al., 2020).

CMSS is evaluated on Digits, DomainNet, PACS, and Office-Caltech10. Reported averages are 90.8% on Digits, 46.5% on DomainNet, 89.5% on PACS, and 97.2% on Office-Caltech10, outperforming the listed baselines on these benchmarks. On the difficult Quickdraw target in DomainNet, it is the only method that improves over Source Only. Qualitative analyses show domain preference over time, sample-wise ranking, weight dynamics consistent with a curriculum, and t-SNE trajectories in which target features move toward different source modes over training (Yang et al., 2020).

DMSN extends the same selectivity principle to object detection under multi-source domain adaptation (Yao et al., 2021). Built on Faster R-CNN, it uses a shared low-level backbone $$\left(t_{\text{trg}}, \{(f^{1}_{\text{src}}, t^{1}_{\text{src}}), \ldots, (f^{k}_{\text{src}}, t^{k}_{\text{src}})\}\right) \mapsto f_{\text{trg}}.$$7, multiple source-specific high-level subnets, and a pseudo target subnet. Low-level features are strongly aligned across all domains, while high-level features are weakly aligned between each source and target, reflecting the paper’s central trade-off between domain invariance and discriminative power. In phase 2, pseudo subnet learning updates the pseudo target subnet by an EMA of a weighted combination of source subnet parameters: $$\left(t_{\text{trg}}, \{(f^{1}_{\text{src}}, t^{1}_{\text{src}}), \ldots, (f^{k}_{\text{src}}, t^{k}_{\text{src}})\}\right) \mapsto f_{\text{trg}}.$$8 where the weights are derived from recent source-target discriminator losses (Yao et al., 2021).

On cross-camera adaptation from Cityscapes + KITTI to BDD100k daytime, DMSN = 49.2 AP, compared with 46.5 AP for the best source-combined DA baseline and 44.1 AP for M$$\left(t_{\text{trg}}, \{(f^{1}_{\text{src}}, t^{1}_{\text{src}}), \ldots, (f^{k}_{\text{src}}, t^{k}_{\text{src}})\}\right) \mapsto f_{\text{trg}}.$$9SDA. On cross-time adaptation from BDD100k daytime + night to dawn/dusk, DMSN = 35.0 mAP, outperforming the best source-combined result by 4.1 mAP and the best multi-source DA baseline by 7.4 mAP. The ablation sequence—28.0, 29.9, 31.7, 32.9, 34.2, 35.0 mAP—supports the contributions of low-level alignment, high-level alignment, pseudo subnet learning, consistency regularization, and dynamic weighting (Yao et al., 2021).

Across CMSS and DMSN, a consistent conclusion emerges: multi-source adaptation is not improved by indiscriminate aggregation. It improves when the system learns which samples, feature levels, or parameter blocks should be emphasized, delayed, merged, or left separate.

The literature therefore treats “multi-source morsels” not as a single algorithmic object but as a family resemblance across problems. Sometimes the morsel is a paradigm cell, sometimes a rank-one update, sometimes a frontier work item, sometimes a source-grounded answer pair, and sometimes a biased memory projection. What unifies these cases is the insistence that source heterogeneity must remain computationally visible. Systems that preserve and manipulate that granularity can be more accurate, more scalable, more robust to perturbation, and more diagnostically transparent than systems that collapse all sources into one undifferentiated aggregate (Kann et al., 2016, Osial et al., 26 Aug 2025, Chakraborty et al., 26 Aug 2025, Li et al., 27 May 2026, Yang et al., 28 May 2026).