Ability-Guided Transfer in AI

• Ability-guided Transfer is a framework that leverages explicit estimates of competence—using latent variables, procedural units, or parameter modulation—to condition transfer decisions.
• It employs diverse mechanisms such as curriculum generation, soft attention routing, and per-parameter guidance to match task difficulty with learner ability and avoid negative interference.
• Empirical evidence in domains like reinforcement learning, complex QA, and multilingual modeling demonstrates improved efficiency, adaptability, and robustness through ability-conditioned transfer.

Searching arXiv for papers on ability-guided transfer and closely related formulations. Ability-guided Transfer denotes a family of transfer mechanisms in which an explicit estimate of “ability” governs transfer decisions. In the literature, the term is operationalized in several non-equivalent ways: as a learner’s latent ability matched to task difficulty in reinforcement learning curricula, as a reasoning ability encoded in decomposition traces for in-context learning, as a per-state usefulness signal over source policies or value functions, as per-parameter plasticity during fine-tuning, as language-agnostic ability weights in multilingual LLMs, and as trace-grounded procedural units in agent self-evolution (Tio et al., 2023, V et al., 2023, Rajendran et al., 2015, Nikolić et al., 2023, Chen et al., 2024, Gao et al., 6 Jul 2026). The unifying idea is not a single algorithmic template, but a design principle: transfer is conditioned on an estimated competence structure rather than applied uniformly.

1. Conceptual scope and recurring structure

Across the cited work, Ability-guided Transfer is characterized by three recurring components. First, an ability representation is defined: a latent scalar $a$ for an RL student, a dataset-level reasoning characteristic $z_s$, a state-conditional attention weight $w_i(s)$, a guidance parameter $g_i$ for each model parameter, an ability-specific weight delta $R(A_i)$, or a canonical procedural unit $u$ extracted from traces (Tio et al., 2023, V et al., 2023, Rajendran et al., 2015, Nikolić et al., 2023, Chen et al., 2024, Gao et al., 6 Jul 2026). Second, transfer is routed through this representation: curriculum generation matches difficulty to ability, exemplars are selected to induce a reasoning ability, source experts are mixed according to state, parameter updates are scaled according to learned plasticity, or only selected tensors/channels are transferred (Tio et al., 2023, V et al., 2023, Rajendran et al., 2015, Nikolić et al., 2023, Chen et al., 2024, Jin et al., 14 Jan 2026). Third, effectiveness is assessed not merely by raw task performance, but by whether transfer avoids negative interference while preserving or recovering desired competencies (Rajendran et al., 2015, Chen et al., 2024, Jin et al., 14 Jan 2026).

Work	Domain	Ability representation
PERM	RL curricula	latent agent ability $a$, environment difficulty $d$
ICAT	Complex QA	dataset-level QD or RG ability
A2T	RL transfer	state-dependent usefulness weights $w_i(s)$
Guided Transfer Learning	neural adaptation	per-parameter guidance $g_i$
MAET	multilingual LLMs	ability-specific weights $z_s$0
EvoAgentBench	agents	canonical trace-grounded Ability units

This breadth matters because the term is not standardized. In some papers, “ability” is an intrinsic property of a learner; in others, it is a property of data, procedures, modules, or parameter deltas. A plausible implication is that the central research problem is less the semantics of the word “ability” than the construction of a transferable control signal that can mediate selective reuse.

2. Formalizations of ability

A prominent formalization appears in the Parameterized Environment Response Model (PERM), which models learner ability and environment difficulty as latent variables inspired by Item Response Theory. For binary success, the paper gives the 1PL model

$z_s$1

and PERM uses a 1PL-inspired continuous-response formulation over reward $z_s$2, with amortized posteriors $z_s$3 and $z_s$4, and decoders $z_s$5 and $z_s$6. Its objective is

$z_s$7

with $z_s$8 (Tio et al., 2023). Here ability is a latent variable on a calibrated scale shared across students.

In ICAT, ability is not a latent learner trait but a dataset characteristic demonstrated through reasoning traces. Question decomposition is represented as

$z_s$9

and transfer is successful when exemplars $w_i(s)$0 from a source dataset induce a faithful decomposition $w_i(s)$1 and answer $w_i(s)$2 on a target dataset (V et al., 2023). The representation is procedural and externalized in prompt exemplars rather than learned in model weights.

A2T defines ability operationally as the per-state usefulness of a source solution. If source policies $w_i(s)$3 and a base policy $w_i(s)$4 are mixed by soft attention weights $w_i(s)$5, then

$w_i(s)$6

and similarly for value functions,

$w_i(s)$7

The attention weights

$w_i(s)$8

realize a state-conditional competence estimator over source and base solutions (Rajendran et al., 2015).

In guided transfer learning, ability is mapped to per-parameter adaptability. Each weight or bias has a guidance value, and updates are scaled elementwise: $w_i(s)$9 At scalar level,

$g_i$0

The guidance values are computed from mean squared or mean absolute deviation across scout models trained on related tasks, then normalized to obtain $g_i$1 (Nikolić et al., 2023).

MAET formalizes ability as a language-agnostic parameter delta. The extracted ability-specific representation is

$g_i$2

explicitly subtracting language-only changes from ability-plus-language changes (Chen et al., 2024). ACT, by contrast, localizes ability to sparse output channels using activation differences

$g_i$3

averaged over answer tokens and samples to produce $g_i$4, then transfers only masked parameter rows associated with the top channels (Jin et al., 14 Jan 2026).

EvoAgentBench adopts yet another formalization. An Ability card is a tuple

$g_i$5

where $g_i$6 is a trigger condition, $g_i$7 a procedure, $g_i$8 supporting evidence, $g_i$9 an applicability boundary, and $R(A_i)$0 (Gao et al., 6 Jul 2026). Ability here is a reusable operational unit, neither latent nor parametric.

3. Transfer mechanisms

In curriculum generation, PERM performs transfer by inferring a new student’s ability from a few interactions and then sampling task parameters from a difficulty-conditioned generator. The policy principle is to set target difficulty near the current ability estimate, typically $R(A_i)$1, then sample $R(A_i)$2 (Tio et al., 2023). Because PERM is trained offline and does not use regret or value-function-derived surrogate signals, it can be reused across students in the same parameterized environment family.

In prompt-based QA, ICAT transfers ability entirely through exemplar construction. It identifies transfer datasets with the needed reasoning ability, selects exemplars, and assembles prompts that induce decomposition or rationale generation in a single inference pass (V et al., 2023). Its automated selector is the Frechet Term Embedding Distance,

$R(A_i)$3

computed from token-embedding Gaussian statistics. Exemplars are chosen by minimizing FTD to the target question, optionally with a diversity term (V et al., 2023).

In A2T, transfer is executed through soft routing. Frozen source experts remain unchanged, the attention network learns per-state weights from returns or TD signals, and the base network learns concurrently so that the system can fall back to target-task learning when all sources are poor (Rajendran et al., 2015). This architecture is explicitly designed to avoid negative transfer and to enable selective transfer across different regions of the state space.

Guided transfer learning uses an update-modulation mechanism rather than architectural routing. Scout tasks determine which parameters have historically moved under related adaptations, and only those parameters receive substantial updates during target training (Nikolić et al., 2023). This reduces the effective search dimensionality and, according to the paper, can help avoid local minima and overfitting under data scarcity.

MAET transfers ability by sparse continual pre-training on key neurons, tensor selection based on similarity to multilingual linguistic deltas, and a training-free merge. For each tensor $R(A_i)$4,

$R(A_i)$5

so ability weights are injected only into tensors deemed least aligned with general linguistic variation (Chen et al., 2024). ACT uses a related but more localized mechanism: masked task-vector addition over selected channels, followed by lightweight fine-tuning for compatibility (Jin et al., 14 Jan 2026).

EvoAgentBench does not itself propose a transfer algorithm, but it decomposes transfer into extraction, routing, and uptake. Methods such as Memento, ReasoningBank, and GEPA differ in how they encode training experience and deliver it at test time, while Anchor Skill serves as a diagnostic upper bound using curated Ability-grounded skills routed by curator-side labels (Gao et al., 6 Jul 2026). This makes explicit that ability-guided transfer can fail at multiple stages even when transferable content exists.

4. Empirical evidence

PERM reports strong representation quality after offline training. On LunarLander, response MSE is $R(A_i)$6, $R(A_i)$7 MSE is $R(A_i)$8, and $R(A_i)$9 of reward regressed on $u$0 is $u$1; on BipedalWalker, response MSE is $u$2, $u$3 MSE is $u$4, and $u$5 (Tio et al., 2023). In LunarLander training, PERM-Online and PERM-Offline outperform PLR⊥, PAIRED, and Domain Randomization during training and on fixed evaluation environments, and episode length increases from about $u$6 timesteps under DR to about $u$7 for PERM agents. The paper also states that PERM-Offline generalizes to new students without loss of training quality on LunarLander (Tio et al., 2023).

ICAT reports gains across numerical reasoning, compositional QA, and heterogeneous QA. Representative comparisons against Few-shot-COT include MuSiQue $u$8 vs $u$9, 2WikiMultiHopQA $a$0 vs $a$1, MultiArith $a$2 vs $a$3, SvAmp $a$4 vs $a$5, and FinQA $a$6 vs $a$7 for ICAT+FTD (V et al., 2023). Static transfer-source choice also matters: FinQA with TabMwp yields $a$8 compared with $a$9 when using AquaRat, and WQA benefits from combining SubQA and StrategyQA rather than using either alone (V et al., 2023).

A2T shows selective transfer in both synthetic and Atari settings. In Pong with two blurred experts, the attention-only mixture achieves an average score of about $d$0 after a single epoch of $d$1 frames, and in Breakout, selective transfer over two blurred experts reaches about $d$2 after $d$3 epochs, compared with individual experts at about $d$4 and $d$5 (Rajendran et al., 2015). The paper also reports that when negative sources such as Inverse-Pong or Freeway are included, A2T suppresses them and performs comparably to learning from scratch rather than suffering the severe degradation observed with naive fine-tuning.

Guided transfer learning reports several low-data and hard-optimization results. On Omniglot $d$6-way one-shot evaluation, guided transfer learned from MNIST pre-training plus $d$7 cousin $d$8-class scouts consistently outperforms pre-training alone; on generalized XOR-like random mapping tasks, guidance avoids long stagnation before breakthrough; and in sequential one-shot transfer, guided transfer yields stepwise accuracy improvements across iterations, unlike pre-training alone (Nikolić et al., 2023).

MAET reports multilingual gains on both math and science. On MGSM with LLaMA-3 $d$9B, average accuracy rises from $w_i(s)$0 to $w_i(s)$1, with ES $w_i(s)$2, ZH $w_i(s)$3, BN $w_i(s)$4, and TE $w_i(s)$5. On multilingual MMLU science tasks, average accuracy rises from $w_i(s)$6 to $w_i(s)$7 (Chen et al., 2024). The paper also reports that MAET preserves or slightly improves out-of-domain performance relative to the base model while continual pre-training baselines degrade on MMLU, HumanEval, MBPP, and OpenBookQA.

ACT reports that in Qwen2.5-7B recovery, transfer-only with top-$w_i(s)$8 per-ability masks and a $w_i(s)$9 union of parameters preserves math while improving science to $g_i$0, above TA, TIES, and DARE at $g_i$1, $g_i$2, and $g_i$3. With fine-tuning, top-$g_i$4 masks and an $g_i$5 union raise science to $g_i$6, compared with $g_i$7 for SFT-only (Jin et al., 14 Jan 2026). The paper further states that ability-related activation differences are typically concentrated in less than $g_i$8 of channels.

EvoAgentBench contributes a different kind of empirical result. On its $g_i$9 train/test split across four domains, curated Ability content in Anchor Skill yields positive average transfer across all three evaluation backbones: $z_s$00 for Qwen3.5-27B, $z_s$01 for Qwen3.5-397B, and $z_s$02 for Gemma-4-31B, whereas no automatic method sustains positive gain in all settings (Gao et al., 6 Jul 2026). This suggests that the bottleneck is not the absence of transferable procedures, but the reliability of automatic extraction and routing.

5. Common challenges and limitations

A recurring limitation is that transfer quality depends on how tightly the control variable tracks realized task difficulty or ability. PERM performs best in deterministic environments such as LunarLander; in BipedalWalker, where teacher-controlled parameters specify ranges rather than exact instantiated levels, the benefits are muted because target difficulty is diluted by stochastic level generation (Tio et al., 2023). This is a general issue for any method whose transfer decision is made in a latent control space but whose execution occurs through a noisy environment.

Another recurrent issue is source-target relatedness. ICAT assumes that transferable reasoning structure exists in the source datasets; performance drops when the source ability mismatches the target, as illustrated by weaker FinQA transfer from AquaRat relative to TabMwp (V et al., 2023). MAET similarly requires that the target ability be learnable from a high-resource language and that language and ability effects can be cleanly decomposed; if tensor or neuron selection is noisy, negative transfer can occur, especially in low-resource languages (Chen et al., 2024). Guided transfer learning also assumes that scout tasks are sufficiently related to the target task; poor scout design can misguide updates (Nikolić et al., 2023).

Selective routing methods face their own difficulties. A2T scales linearly with the number of source networks because every source must be evaluated at each state, and overconfident gating can overweight suboptimal experts (Rajendran et al., 2015). EvoAgentBench demonstrates that even when the transfer unit is operationally well-defined, automatic methods may fail at extraction, routing, or uptake, with catastrophic negatives such as $z_s$03 for Memento on Nanobot / Qwen3.5-27B / SWE (Gao et al., 6 Jul 2026).

There are also representational assumptions. PERM’s experiments primarily rely on 1PL-like scalar latents, though the paper notes that real-world skills may be multidimensional and that vector latents with 2PL or 3PL analogs could improve fidelity (Tio et al., 2023). ACT assumes shared architecture and tokenizer across models, and MAET is evaluated only on math and science with $z_s$04B-scale backbones (Chen et al., 2024, Jin et al., 14 Jan 2026). These constraints indicate that ability-guided transfer is often effective only within a carefully delimited transfer family.

6. Relation to adjacent paradigms and broader significance

Ability-guided Transfer intersects with, but is not reducible to, several established paradigms. In RL, A2T is explicitly connected to mixture-of-experts and options frameworks, but its distinctive feature is soft, differentiable, state-dependent routing over frozen sources plus a trainable base expert (Rajendran et al., 2015). PERM connects curriculum learning to Item Response Theory and the zone of proximal development by choosing tasks near the learner’s inferred capability (Tio et al., 2023). In NLP, ICAT differs from standard Chain-of-Thought prompting because it reuses verified expert decompositions or rationales from other datasets rather than creating new task-specific demonstrations (V et al., 2023).

In model adaptation, guided transfer learning resembles per-parameter learning-rate modulation, while MAET and ACT relate to task-vector and model-merging methods. What distinguishes them is the explicit attempt to localize ability-specific substructures before transfer: scout-derived plasticity in one case, key neurons and tensor filtering in another, and activation-difference channel localization in the third (Nikolić et al., 2023, Chen et al., 2024, Jin et al., 14 Jan 2026). This suggests a broader shift from global transfer operators toward targeted transfer mechanisms that try to isolate where competency resides.

EvoAgentBench extends the idea beyond model internals to agent trajectories. Its Ability Graphs make procedural overlap explicit and separate transfer into experience encoding, routing, and uptake (Gao et al., 6 Jul 2026). A plausible implication is that future work may increasingly treat abilities as compositional operational objects rather than opaque properties of a monolithic model.

Taken together, these works portray Ability-guided Transfer as a general research program rather than a single technique. Its central claim is that transfer improves when mediated by an explicit competence signal—latent, procedural, architectural, or parametric—that constrains what is reused and where it is applied. The empirical record summarized here supports that claim in several settings, while also showing that reliable ability estimation, compatibility of source and target structure, and robust routing remain the principal unsolved problems (Tio et al., 2023, V et al., 2023, Rajendran et al., 2015, Chen et al., 2024, Gao et al., 6 Jul 2026).