PyTorchKR Papers

최신 논문

총 523,801편
태그·기간 등 상세 필터

카테고리 탐색

arXiv:2611.05166v1cs.CL

When Self-Verification Hurts: Calibrated Answer Repair for Financial Question Answering

Financial question answering can fail even when a model extracts a plausible number from the supplied evidence: the answer may attach it to the wrong entity, period, unit, or accounting concept. Asking the model to verify and revise its draft can help, but open-ended revision may alter content that was already correct or inspect the wrong relation and reinforce the original error. We formulate repair as a decision that balances the information it may recover against the correct content it may damage. Based on this trade-off, we introduce Calibrated Answer Repair (CAR), which first estimates whether a draft is likely to benefit from repair, then selects a targeted check, localizes the relevant evidence or draft span, and limits the permitted edit. Across FinRank, Fin-RATE, and FinQA with three open-weight models, CAR outperforms generic verification in all dataset-model settings. At the system level, it recovers 14.6% of initially missed answer requirements while damaging 2.1% of initially correct ones. Its post-draft token cost is over 30 percent lower than the iterative revision alternatives. Ablations on two dataset-model settings show that evidence localization and bounded editing are the main contributors to CAR's gains.

arXiv:2611.05164v1cs.LG

Basin2Vec: Learning Basin-Centric Embeddings for Scalable Hydrological Modeling

Reliable streamflow prediction across time and in ungauged basins requires representations that capture hydrologically relevant basin characteristics. Existing methods rely on either fixed handcrafted attributes or general-purpose Earth observation embeddings not tailored to basin structure. We introduce \textbf{Basin2Vec}, a self-supervised framework that learns compact basin representations from hydroclimatic and physiographic data without observed streamflow. Basin2Vec organizes representation learning around basin scale and relationships across space and time. Across multiple forecasting architectures, it consistently outperforms predefined attributes and general-purpose geospatial embeddings in temporal generalization and zero-shot spatial transfer. Basin2Vec provides average relative gains of up to 8.3% temporally and achieves 3.5x the median predictive performance of the strongest baselines on unseen basins. These results demonstrate the effectiveness of basin-centric pretraining for transferable hydrologic prediction. Code is available at https://github.com/Muhammad-Talha-MT/basin2vec

arXiv:2611.05158v1cs.CV

Perturb, Repair, Verify: Self-Play Vision-Language Verifiers for Compositional Understanding

Vision-language verifiers judge whether an image faithfully matches a textual description, yet existing training pipelines typically rely on static supervision from stronger external judges or curated hard-negative data. Such supervision requires external generation or annotation, is teacher-bounded, and is difficult to adapt as model failures evolve. We introduce PROVE (Perturb-and-Repair Optimization for Vision-Language Evaluation), a self-play framework that trains verifiers through a closed perturb-verify-repair-verify loop. Starting from a known-correct image-caption pair, a Mutator introduces a typed perturbation over objects, attributes, counts, or relations. The verifier then judges the perturbed caption and provides textual feedback, which a Repairer uses to recover a faithful caption for re-verification. These two verification rounds produce structured rewards for informative mismatches and visually grounded repairs. Since perturbation, verification, and repair are role-conditioned behaviors of the same evolving multimodal backbone, updating the shared parameters also improves the verifier itself. Experiments show that PROVE improves fine-grained compositional verification across model scales and families and strengthens visual-detail robustness on the reference backbone, with limited, benchmark-dependent changes on broader multimodal tasks.

arXiv:2611.05157v1cs.LG

Macrocanonical Generator Networks: Data-Efficient Neural Surrogates for Amortized Physics Simulation

Many scientific simulations are computationally expensive, limiting their use in simulation-based inference, uncertainty quantification, and decision making. We introduce macrocanonical generator networks, a framework for learning fast, data-efficient neural surrogates for amortized simulation of stationary physical processes by matching multiscale statistics of a target process. Inspired by microcanonical maximum-entropy synthesis, which matches feature statistics per sample by inner-loop optimization, our approach instead trains a neural generator to satisfy the same constraints in expectation, preserving physically realistic sample-to-sample fluctuations. We prove that gradient-descent training induces preconditioned gradient-descent dynamics in sample space that inherit the symmetry-preservation properties of the microcanonical gradient descent of Bruna and Mallat, amortizing its per-sample inner-loop optimization into a single training run after which every sample is generated in one forward pass. We further show that residual generator parameterizations initialized near the identity, trained with a small learning rate, preserve entropy locally during early optimization and mitigate mode collapse. Experiments on cosmology and fluid simulations show that macrocanonical generators recover realistic samples from extremely small training sets, in some cases a single training example, while generating new realizations more than $10^6$ times faster per sample than the reference simulator and more than $10^3$ times faster than microcanonical gradient descent. The trained generators can be reused for downstream pretraining, including simulation-based inference, providing a data-efficient route to multi-fidelity scientific machine learning workflows.

arXiv:2611.05155v1cs.RO

Asynchronous Robot Policies

Scaling vision-language foundation models has significantly advanced generalist robot manipulation, but the increased model scale introduces computational latency that degrades real-time physical reactivity. Existing Vision-Language-Action (VLA) architectures execute high-level semantic reasoning and action ex- ecution sequentially within the primary control loop, creating a fundamental tension between semantic-reasoning capability and robot agility. To address this, we propose Asynchronous Robot Policy (ARP), an architectural framework that decouples semantic reasoning from low-level action generation using asynchronous processes operating at independent frequencies. In ARP, a high-capacity Vision-Language Model (VLM) asynchronously updates a semantic memory at low fre- quency, while a lightweight Vision-Action Expert, conditioned on this semantic memory, generates robot actions at a much higher control frequency. Such a formulation, enables having a fast perception-action loop without compromising on the semantic capability. To overcome asynchronous memory staleness, ARP trains the vision-action expert on the semantic staleness and execution delay, with temporally mismatched observation across processes. On Franka manipulation tasks, ARP achieves an 18% higher success rate and 20% less time need for execution in quasi-static setups, while 3-5 times baseline performance in dynamic simulation environments. Furthermore, ARP enables operating the VLM at 10 times lower than its nominal frequency with no significant performance loss making it much more efficient than the baseline. These results highlight the architectural advantages of modeling robot policies as asynchronous processes, enabling real-world reactive deployment without sacrificing the scale of foundation models.

arXiv:2611.05151v1cs.AI

Catalyzing Interactive HTML Generation Capabilities via High-Fidelity Reinforcement Learning

Generating HTML from natural-language specifications is an important application of code large language models (LLMs). Despite producing renderable pages, current models often fail to ensure functionality and visual quality during actual interaction. Existing methods typically target task-specific webpage supervision, leaving demanding interactive artifact optimization underexplored. Using mini-games as a challenging testbed, we expose limitations in interaction experience, reward signal, and credit assignment fidelity that are obscured by shallow webpage interactions. We introduce WebCatalyst, a browser-interaction reinforcement learning framework with three complementary components: a game-level GUI assessor for sustained interaction and evaluation, Visual Anchor for mitigating VLM reward hacking, and Action-Trace Credit for grounding token-level credit assignment in runtime execution traces. Trained on fewer than 1K requests, predominantly for games, WebCatalyst raises the Cookie-Bench Mini Games score from 68.8 to 75.3 and also the WebGen-Bench accuracy from 48.8 to 56.9 on Qwen3.6-35B-A3B, outperforming Claude Opus 4.7 in average rank across evaluated metrics. Further analysis shows that Visual Anchor prevents aesthetic degradation, while runtime-grounded credit assignment accelerates functional improvement. Together, our results show that improving feedback fidelity is the key to substantially advancing interactive HTML generation.

arXiv:2611.05148v1cs.AI

Creating Power Flexibility in Datacenters via Query Routing

The growing energy requirements of large language model inference highlight the need for datacenters to support safe grid operation by providing power demand flexibility, while maintaining service for interactive workloads. This work investigates query routing as a mechanism to enable such capability by dynamically allocating queries across different language model instances without deferring requests. We develop a general analytical framework that connects query arrival rates and sequence length distributions to token iteration time, request residence time, serving capacity, and power consumption of model instances. Leveraging this framework, we characterize the feasible query routing allocations and the corresponding power demand flexibilities, enabling the selection of routing strategies to achieve desired power consumption behavior. We validate our framework across four GPU platforms on language models ranging from 1B to 70B parameters, obtaining 2.8% mean absolute percentage error in estimating achievable power flexibility. We evaluate our approach on a real trace-driven simulation of a 10,000 H100 GPUs datacenter, exhibiting flexibility up to 16% of total power consumption, with 90% of the desired power reduction achieved within approximately 10 seconds. These results establish query routing as an effective mechanism for datacenter power demand flexibility and demonstrate the broader applicability of our proposed framework to datacenter level planning and orchestration.

arXiv:2611.05147v1cs.CL

GM League Arena: An Interactive Long-Horizon Task Benchmark Based on Professional Sports League Management

Long-horizon decision-making under sustained competition requires large language models (LLMs) to balance immediate investment with future resource needs. However, local success in multi-step interactions can not adequately capture cumulative decision effects or demonstrate long-term goal achievement under resource constraints and sustained competition. To address this gap, we introduce GM Bench, a benchmark for managing professional sports teams across seasons under resource constraints and competition among teams. We build the environment simulators for four sports and add financial constraints, player demands, and market mechanisms to establish a unified task across sports. Diverse rosters of comparable strength are constructed from real-world player data, and each LLM is evaluated as a team general manager to assess its long-term management capabilities. We evaluate competitive performance, cash reserves, and cross-season improvement across four sports. Over five NBA seasons, Claude Opus 5.5 ranks first among 29 LLMs, while GPT-6 Astra and GPT-6.1 Sol combine strong performance with larger cash reserves.

arXiv:2611.05145v1cs.CV

When Cell Crops Share a Parent: Measuring and Isolating Acquisition Leakage in Cervical Cytology Benchmarks

Public cervical cytology benchmarks release many cell crops cut from far fewer microscope fields or slides and are often split at the crop level, so test cells share acquisition parents with training cells. In a preregistered study, crop-level splitting raised SIPaKMeD accuracy by 0.97 percentage points across twelve architectures and twenty seeds (95% CI 0.59-1.39), and CRIC accuracy by 12.08 points (11.16-12.99). On SIPaKMeD, a factorial that fixes training crop and parent counts attributes the gain to sibling exposure (1.73 points, 0.95-2.51), not to the extra parent diversity of crop-level splits (0.00). Parent identity is decodable at about 100 times chance, yet erasing 82% of its above-floor linear recoverability left the gap unchanged, and Grad-CAM did not distinguish leaking from clean models; matching test cells' color statistics to an unseen same-class field removed most of the advantage. Decodability and saliency cannot substitute for a parent-disjoint split.

arXiv:2611.05141v1cs.LG

Beyond Foundation Models: PFN Priors Improve Tabular Hyperparameter Optimization

Synthetic priors have so far served one purpose: pretraining tabular foundation models. We ask whether the same priors can improve conventional tabular models, which are trained from scratch on each dataset. We propose PriorTune, a method that meta-tunes hyperparameter configurations on synthetic tasks alone. Offline, for a given model, we fit a pool of configurations on datasets sampled from a published PFN prior and greedily select the subset with the highest winrate over the random baseline. Online, a new dataset fits only the selected configurations, so the gain comes at no extra cost. Empirically, we observe that scaling both the number of configurations and synthetic datasets has an equal scaling effect on the performance of the final subset of configurations, which we also explain theoretically. On the full BeyondArena benchmark, LightGBM with TabPrep features with our configurations ranks above TabPFN-3 and RealMLP, while being much more efficient at inference. Overall, our results establish existing PFN priors as a scalable source of meta-training data for hyperparameter selection for non-foundational tabular models.

arXiv:2611.05139v1cs.AI

Meta$^2$: Precomputing Reusable Knowledge for Online Harness Evolution in Long-Horizon Agents

Harness optimization improves long-horizon agents by refining the prompts, tools, memory, and hooks around a frozen model. Offline methods can learn from completed trajectories, but deploy one shared harness that may overfit training tasks and cannot react to evidence revealed during a new run. Online evolution can adapt before, during, or after one run through an evolver, but may have to reason about every intervention from scratch or rely on verifier feedback that is usually unavailable at test time. We introduce Meta$^2$, which moves offline optimization one level up: rather than optimizing a shared harness, it learns reusable knowledge for harness evolution and applies it within each task online. Offline, Meta$^2$ learns a textual Meta-Skill $M$ by contrasting training rollouts of the same task agent under three configurations: without an evolver, with an evolver guided by the current version of $M$, and with an evolver given privileged verifier information available only in training. Contrasts in trajectories and outcomes are aggregated across tasks to refine $M$ into reusable guidance on when and how to evolve, when to leave the harness unchanged, and which interventions to avoid. Online, an evolver combines the frozen $M$ with emerging trajectory evidence to decide whether to KEEP the harness unchanged or EVOLVE it through a task-local edit. Each edit applies to the rest of the same run without resetting its environment or progress. Thus, evolution guidance transfers across tasks while harness edits remain task-local, and held-out verifier feedback is never available to the evolver. On two long-horizon benchmarks, Terminal-Bench 2.1 and DeepSWE 1.1, Meta$^2$ achieves the highest performance among evaluated methods with GPT-5.6-Terra, reaching 81.00% and 44.44%, respectively. Ablations support the value of contrastive evolution guidance and keeping online harness edits task-local.

arXiv:2611.05137v1stat.ML

A Random Matrix Analysis of Symmetric Tensor Unfolding

We study a fundamental problem in machine learning: partial recovery of a planted signal $\boldsymbol{v} \in \mathbb{S}^{n-1}$ from a spiked tensor $\boldsymbol{\mathsf{X}} = \lambda \, \boldsymbol{v}^{\otimes p} + \boldsymbol{\mathsf{Z}}$, where $\boldsymbol{\mathsf{Z}}$ has symmetric Gaussian entries of variance $\asymp n^{-1}$. While previous works replace $\boldsymbol{\mathsf{Z}}$ by an i.i.d. standard Gaussian tensor to circumvent technical difficulties caused by dependencies, this paper directly analyzes the symmetric model via random matrix theory. We show that the square unfolding algorithm is subject to a sharp phase transition for partial recovery at $\lambda \asymp n^{(p-2) / 4}$, validating an open conjecture from Montanari and Richard (2014). We get precise expressions for the limiting spectral distribution, deterministic equivalent, phase transition threshold and overlap. The main message is that unfolded symmetric tensors spectrally behave like standard Wigner matrices, up to simple transformations accounting for high-order symmetries. We consider extensions to multiple orthogonal spikes, recursive unfolding and non-square unfolding via disproportionate random matrix theory, leading to an interesting conjectured statistical-computational tradeoff regarding the choice of unfolding: square versus long.

arXiv:2611.05135v1cs.AI

From Opinions to Commitments: Unstable Assignments and Silent Choices in AI-Mediated Proposals

Large language models (LLMs) are increasingly being used to turn citizens' opinions into collective proposals. These proposals can make assignments: who should act, what must be done, how much one should spend and by when. However, if the assignments change when the same opinions are processed again, publishing a single proposal can hide alternatives that participants never get to see. We examined this issue in four LLM mediators (three open-weight and one commercial) using opinions from DeliberationBank and Polis. For each task, a mediator drafted a proposal from thirty opinions five times, varying either the random seed or the order of the opinions. In 17% of five-run sets from the open-weight mediators and 20% from the commercial one, some runs contained assignments while others did not. Reordering the opinions changed the assignments almost as often as replacing them with opinions from different citizens. Three human annotators confirmed that the detected differences between runs were real, although the LLM judge overstated how often they occurred. Substantial variation remained under four safeguards: temperature zero, a commercial mediator, draft selection, and draft merging. The mediator's objective also shaped representation: prompts seeking agreement covered minority positions much less often than prompts asking to describe differences. A single proposal can therefore obscure choices in both the assignments it proposes and the views it excludes. We propose audits that make alternative assignments visible and disclose the objective behind each proposal.

arXiv:2611.05130v1math.AP

Solving the Nonlinear Westervelt Equation in Ultrasound Propagation Using Physics-Informed Neural Networks

Modeling the nonlinear acoustic wave propagation in biological tissues needs accuracy, especially in ultrasound imaging and therapeutic applications. The 1D Westervelt equation, \( p_{xx} - p_{tt} + \gamma\, p_{ttt} + \eta\,(p^{2})_{tt}=0, \) could be used to describe a finite-amplitude ultrasound propagation in these tissues. This equation accounts for nonlinearity, \(\eta\), and attenuation, \(\gamma\). To solve this equation, we observe that traditional solvers may struggle to do so correctly. This is due to the difficulties arising from the nonlinear nature of the equation. We solve the equation by defining three synthetic cases - linear (\(\eta=0.001\), \(\gamma=10^{-6}\)), shock-dominated (\(\eta=0.02\), \(\gamma=10^{-6}\)), and dissipation-dominated (\(\eta=0.01\), \(\gamma=5\times10^{-4}\)) - and using Physics-Informed Neural Networks (PINNs) to showcase the solution. A sinusoidal activation function is used to capture the oscillatory behavior of the problem. The results of our study, with PDE residual converging to at least \(10^{-6}\), boundary condition loss to \(10^{-5}\), and the PINN loss converging to \(10^{-5}\) for each case at \(30,000\) epochs, show that \(\gamma\) scales the rate of amplitude decay, while \(\eta\) governs frequency-dependent phase speeds.

arXiv:2611.05128v1cs.LG

Do Ray-Traced Cities Make a Better Channel Prior for an Unseen Site? A Pre-Registered Study in Simulation, and a Test on Measured Channels that Reverses It

Channel estimators at a newly deployed base station usually start from 3GPP stochastic channel models, which describe no particular street. We ask whether a channel prior pooled from many ray-traced cities serves an unseen site better than the 3GPP CDL models, and whether learned estimators make the difference. With Sionna RT on 14 city scenes, LMMSE estimation from covariance priors, leave-one-city-out evaluation, and pass/fail bars written down before each main experiment, we find in simulation that the pooled prior sits about 0.6 dB from a site-trained estimator with no site measurements, about what CDL reaches after 50 measurements. Learned models added little: at most 0.65 dB in estimation, and in channel prediction a simple classical method beat CNN predictors by 1.8-12 dB from 2 ms ahead. On measured channels from the DICHASUS outdoor dataset (1.272 GHz, three 2x8 arrays on one building) the result reversed: the pooled ray-traced prior was 3.2-4.1 dB worse than CDL. The simulated channels were too clean; with users at the real distances, even a scattering level tuned from 50 site measurements left the prior about 2 dB behind CDL. A pre-registered follow-on on two more measured environments found the generic CDL prior 2.1-4.7 dB from a site-trained one: no further in a factory hall than on the street, furthest in a small lab room. Finally, a hallway model calibrated to measurements by differentiable ray tracing reproduced delay spread, but as a prior in an unmeasured part of the hallway it was 4.5 dB worse than CDL on the full array. All measured results come from one sounder.

arXiv:2611.05124v1cs.CL

The Right Source Is Not Enough: Understanding History Interference in Speech-LLM Memory

Speech large language models retain audio history to support multi-turn conversations. Yet remembering an event does not ensure that its information is used only when relevant. We investigate why irrelevant historical audio changes answers even when the queried event remains represented. Controlled experiments on MOSS-Audio and MiMo-Audio show that historical values direct errors toward specific competing answers, beyond a generic effect of context length. Causal interventions distinguish two paths. In open-ended question answering, most history-value errors disappear when subsequent computation cannot directly access the irrelevant source. In multiple-choice answering, many errors persist: early historical influence is carried by question-candidate states even after direct source access is blocked. Low-dimensional interventions further separate source-role information from answer-value information. Their effects peak at different depths, and changing the source-role signal does not eliminate the causal influence of competing values. On naturally occurring errors, strengthening the queried role and counteracting the historical value each improve answers; their joint intervention gives the greatest recovery. These findings support a failure of downstream evidence use beyond source identification alone: source-role information can influence a decision without exclusively controlling which historical values affect it. Reliable conversational memory therefore requires not only retention and retrieval, but also control over how retained information enters answer computation.

arXiv:2611.05120v1cs.LG

Towards an Optimal Control Foundation of Early Exits - Turnpikes and ResNets

The feed-forward propagation through neural networks is crucial for training and inference. To trade off efficiency and accuracy, it has been proposed to consider intermediate outputs of hidden states as data-dependent regularization during training and, ultimately, as early-exits during inference. In this work, we analyze the joint optimization of loss functions computed for the output layer and intermediate layers using optimal control theory. We prove that ResNets trained in this joint fashion exhibit the so-called turnpike phenomenon, this implies that the number of layers required to reach a prescribed loss level is bounded from above independently of the overall network depth. This induces a similarity property across the internal representations of early-exiting networks of varying depths. We connect the turnpike to the overthinking phenomenon, and provide theoretical insights into the success of early-exit criteria derived from softmax probabilities, specifically entropy and max-probability. In a range of experiments, we demonstrate the existence and effect of turnpikes and the resulting similarity in the early-exit performance for ResNets of varying depths.

arXiv:2611.05115v1cs.DC

Distributed Linear Bandits for Freshness-Aware Preprocessing in MEC-Enabled Digital Twin Networks

Digital twins (DTs) require timely updates from Internet of Things (IoT) devices to accurately represent the state of the corresponding physical systems, but transmitting large data volumes can compromise freshness. Moreover, different task classes have distinct computational demands and freshness requirements, while stochastic task generation and sizes complicate resource allocation. In this paper, we propose a freshness-aware framework in which multi-access edge computing (MEC) servers preprocess IoT data, reducing transmission volume and DT-server processing load. The framework establishes mutual support between DT and MEC servers. The DTs determine feasible CPU-allocation strategies guaranteeing class-specific age-of-information (AoI) requirements under specified workload, delay, and update-generation bounds. MEC servers collaboratively identify the feasible strategy minimizing expected freshness-normalized preprocessing delay. We formulate CPU-strategy selection as a fixed-confidence distributed best-arm identification problem in linear bandits, with MEC servers as agents and CPU allocations across task classes as arms. We employ DistLinGapE, a distributed collaborative learning method, to coordinate exploration and share observations. Numerical results show that the method identifies the optimal feasible strategy and reduces preprocessing delay when deployed under stationary task-class assignments and workload statistics. Collaboration yields an identification speedup nearly proportional to the number of participating MEC servers in the evaluated settings.

arXiv:2611.05114v1cs.LG

Beyond Multi-Index Models: Learning Intersections of Degree-2 PTFs

We give an algorithm for PAC learning intersections of $k$ degree-2 polynomial threshold functions over $R^d$ under a Gaussian distribution. Our algorithm achieves classification error $\epsilon$ using $d^{O(k)}2^{O_k(\log^3(1/\epsilon))}$ time and samples, improving over the $d^{O(k^2/\epsilon^2)}$ running time of the polynomial regression approach. To our knowledge, our result is the first to obtain a dimension exponent independent of the accuracy parameter, even for $k=2$. We complement our algorithmic result with an SQ lower bound showing that the complexity dependence of $d^{\Omega(k)}$ is unavoidable. Lifting the input to polynomial features turns the target into an intersection of $k$ halfspaces. This suggests using the first two labeled moments for subspace recovery. In the lifted space, however, these moments can reveal nothing. Our starting point is a Gaussian comparison theorem showing that, for sufficiently regular defining quadratics, the labeled moments needed for subspace recovery are approximately preserved. To handle arbitrary defining quadratics, we establish a new structural result showing that every direction with significant resampling influence can be detected through degree-$O(k)$ moments. These results yield an initial $d^{O(k)}2^{\mathrm{poly}_k(1/\epsilon)}$-time learner using adaptive restrictions. To improve the exponential dependence on $1/\epsilon$, we develop a two-stage dimension-reduction approach that avoids the costly adaptive partitioning. After identifying a subspace containing all directions with large influence, we use generalized regularity and Gaussian comparison to recover the remaining relevant quadratic features through a new derivative-based moment operator. The recovered features support an intersection close to the target. Combining them with robust hinge-loss minimization gives our quasipolynomial-time learner.

arXiv:2611.05108v1cs.AI

From AutoResearch to Continual AutoResearch: Do Agents Learn Across Tasks?

Autoresearch has shown that agents can autonomously write code, run experiments, and iteratively improve machine learning solutions. Yet most evaluations start a new agent session from scratch for each new problem, making them costly and leaving unclear whether agents can retain and apply prior insights when adapting to new tasks. We introduce CONTINUAL AUTORESEARCH, an open-ended environment that compares persistent and reset instances of the same agent under matched task streams, data, and time budget. Continual AutoResearch covers nine task families, spanning synthetic tasks, symbolic regression, time-series forecasting, and tabular AutoML. Across five frontier agents, we measure first-solution quality, research efficiency, and final held-out performance. To distinguish insight transfer from session continuity, we run a memory ablation that vary what information agents retain from prior tasks. We show that persistent state improves research efficiency of all five agents, allowing them to find stronger solutions sooner. However, these gains do not consistently translate into better final solutions. Overall, our findings suggest that retaining experience and utilizing it effectively are distinct capabilities.