ColaBench tackles the problem of evaluating how personal agents use long-term interaction history to guide the acquisition and application of external knowledge. The core idea is a traceable benchmark built on an Evidence Graph $\mathcal{G}=(V,E)$ that links personal facts to their disclosures and to supporting documents, paired with MemKB, a retrieval method that connects state summaries and memory-document links. This matters because current benchmarks test memory recall and retrieval in isolation, leaving open whether agents can jointly recover historical conditions and apply newly retrieved rules to a specific user's situation.

Document-level LoRA adapters are pitched as modular knowledge units for retrieval and merging, yet standard open-book evaluation cannot tell whether an adapter has memorized a document or merely learned to read it well. This paper introduces a cloze probe that withholds the source text and finds that open-book scores are almost entirely generic reading behavior: a random adapter recovers 82% of a specialist's open-book advantage but only 2% of its cloze gain. The findings argue that adapter-bank evaluation needs a closed-book probe alongside open-book scores, and the authors' candid disclosure of four silent errors is a model of experimental hygiene.

Natural-language robot control requires mapping user requests to predefined tool subsets without inventing new capabilities. This paper frames tool routing as bounded multi-label selection over a fixed catalog and compares three ChromaDB retrieval policies against a cloud-hosted 120-billion-parameter LLM. The core finding is a clean quality-latency trade-off: clause-aware retrieval (CDBv3) runs locally in milliseconds but lags the LLM in routing accuracy, making it suitable for connectivity-constrained or privacy-sensitive deployments.

Neural networks trained independently on similar tasks often learn compatible semantic structure but express it through incompatible coordinate systems. This doctoral thesis proposes that semantic correspondence—identifying which elements refer to the same meaning across systems—is a fundamental primitive for revealing shared intrinsic geometry. It unifies fourteen methods spanning vision, language, reinforcement learning, and neuroscience under a single framework, distinguishing correspondence-parallel operators (independent semantic projections) from correspondence-coupled operators (direct cross-model maps). The work enables zero-shot transformation and model stitching without joint training, yet explicitly disclaims formal theoretical guarantees, positioning itself as a systematic empirical demonstration.

LAST-V2X addresses the problem of testing how delayed, missing, or corrupted V2X messages compromise safety-critical driving decisions. Its core idea is a framework that translates natural-language test descriptions into executable CARLA scenarios, then causally verifies whether any observed violations are actually due to degraded communication rather than ordinary traffic risk. This matters because existing adversarial methods perturb only the physical world, whereas V2X failures propagate from the information layer and must be isolated through paired counterfactual execution.

PRS tackles automatic heuristic design by making diversification an explicit generation-level operation rather than an emergent property of population management. The framework factorizes heuristic transformation into three semantic directions—expansion, compression, and steering—and rewrites perturbed ideas into executable code via LLMs within a breadth-first search with adaptive operator control. This matters because existing evolutionary frameworks mutate code without directly controlling the semantic trajectory of change, which often leads to premature convergence.

This paper addresses the cold-start forecasting problem for land-restoration sites, where physical models need site-specific hydraulic parameters and traditional data-driven models require months of per-site training data. The core idea is to evaluate TimesFM, a 500M-parameter time-series foundation model, in strict zero-shot mode on 15-minute in-situ telemetry from 27 soil-moisture and temperature sensors across three management regimes. The work matters because it systematically maps where zero-shot foundation models win and lose against classical and tree-based baselines, and quantifies the inference cost on commodity CPU hardware for operational deployment.

Autonomous driving vision–language–action (VLA) models promise explainable decisions by generating natural-language reasoning before predicting trajectories. This paper subjects that promise to a rigorous causal test via counterfactual interventions: the authors perturb model-generated reasoning text and photorealistically remove visual objects that the reasoning cites, measuring whether trajectories actually track these changes. Their findings across four state-of-the-art models indicate that the stated reasoning is weakly coupled to the resulting driving actions, raising serious concerns about the reliability of such explanations in safety-critical deployment.

When agents share a tool catalog and pool their usage experience, adding that usage to the retrieval index boosts tools with usage but can push unused tools out of the top results. This paper shows that folding usage examples into the index severely lowers held-out Hit@5 on two of three catalogs, and proposes Synapse, which scores documentation and usage as separate z-normalized fields before combining them with a learned weight. The finding matters because multi-agent systems risk silently failing on rarely-used tools unless the combination rule explicitly protects them.

Long-text image-text alignment underpins fine-grained scene understanding, yet existing models suffer from saliency collapse: they over-rely on visually prominent text while neglecting informative non-salient context. This paper proposes RoSA, a finetuning framework that decomposes image-text pairs into complementary salient and contextual views and enforces auxiliary contrastive objectives over both. By treating saliency as a multimodal training prior with no inference overhead, RoSA improves retrieval robustness when salient cues are weakened or removed.

Diffusion language models decode many tokens in parallel, making schema enforcement harder than in autoregressive models: two locally valid tokens can be jointly incompatible. DConstrain compiles schemas into context-parameterized transition programs that precompute local grammar steps while leaving the runtime stack and field history as unbound parameters, resolving them only during warm compilation. This factorization avoids the exponential blow-up that forces prior exact decoders to fix property order, yielding large gains in compilation coverage, decoding latency, and task accuracy.

Sparse attention must approximate the long tail of unseen KV-cache tokens, yet discarding it biases the output and random sampling misses heavy hitters at extreme sparsity. OpenAttention folds the tail into a few count-aware mean proxies discovered via query-geometry clustering, achieving 99% of dense accuracy while reading only 1% of the cache. The accompanying Triton kernel delivers a 10× speedup in the decode attention stage, making ultra-low-density attention practical for long-context inference.

RetiJev tackles the latency, cost, and privacy barriers of autoregressive LLMs in point-of-care ophthalmic decision support. The authors introduce RetiJev-Data, a 128k-question source-grounded corpus, and a family of 0.8B–9B compact models that score predefined candidate answers in a single pass instead of generating free-form text. This yields sub-200 ms inference per case on one GPU while reportedly surpassing a 27B LLM and a commercial decision model on guideline grading and clinical-action accuracy.

This paper tackles the challenge of translating natural-language optimization problems into correct, executable solver programs using frozen LLMs. It introduces OptiCycle, a lightweight inference harness that iterates code generation, execution, and repair in a persistent Python state, and OptiState, which adds an explicit five-element revisable formulation. Across nine benchmarks spanning 2,634 problems and 11 Qwen variants, execution-guided revision produces broad accuracy gains and outperforms heavier multi-agent workflows and post-trained specialist models without any parameter updates, though the extra formulation structure helps only some backbones.

Direct Forcing tackles few-step autoregressive text-to-video generation by eliminating the costly ODE initialization stage that prior distillation methods require. The authors propose a single-stage recipe that jointly learns causal continuation alignment and rollout distribution matching from reference videos, using maximum mean discrepancy and dynamic gradient balancing. Starting from a bidirectional Wan2.1-$1.3$B checkpoint, the method trains in just $250$ updates and achieves $84.63$ VBench Total, offering a markedly shorter path to a fast causal generator.

MIG-COW addresses a critical vulnerability in deepfake detection: the difficulty of crafting adversarial examples that simultaneously fool both CNN and ViT architectures. The method combines momentum-integrated gradients with a consensus-orthogonal decomposition to separate shared vulnerabilities from model-specific decision cues. While it achieves near-perfect white-box success, the work is most valuable for highlighting how fragile cross-architecture transferability remains in practice.

VLTest addresses the problem of testing vision-language models (VLMs) where mutating an image, a text prompt, or both can change the expected answer, making naive output-change counting unreliable. Its core idea is to generate controlled visual variants via VQGAN latent code editing and textual variants via a rewrite–expand–prune pipeline, then confirm failures with an automated oracle and attribute joint failures to the specific image-text combination using matched four-cell evidence. This matters because VLMs are increasingly embedded in software systems, and existing robustness evaluations systematically overcount failures by ignoring whether an input still has the same correct answer.

Scientific optimization is expensive and constrained, yet the search policy is usually fixed ahead of time. ADAPT treats the complete policy $\pi=(W,O)$—where to start and how to move—as the object of optimization, using historical designs for target-aware initialization and multi-agent LLM evolution to adapt the procedure itself. Evaluated on eight held-out gravitational-wave detector topologies under matched budgets, it reduces mean loss by 54.9% versus the best classical optimizer and by 62.2% versus the best agentic baseline. The work is compelling within its domain but leaves open questions about generality and reproduction.

LLM agents improve by distilling experience into external knowledge, but different analyses of the same trajectories can produce guidance with wildly different utility. This paper introduces SPLICE, a framework that adapts consistency regularization to knowledge evolution: it perturbs the textual instructions that transform experience into skills or memories, replays tasks under each resulting view, and updates knowledge only where utility gains are consistent across views. A co-evolution loop interleaves this knowledge evolution with GRPO and gated contrastive self-distillation to internalize the guidance. The core idea matters because current agents can be misled by plausible-looking but ineffective skills, and SPLICE offers a principled filter for generalizable knowledge.

This randomized experiment with 297 knowledge workers and 594 essays tests dynamic metacognitive scaffolds for goal specification and task decomposition in AI-assisted writing. The authors show that scaffolds improve output quality, prompting behavior, and three-day topic recall, even after the support is removed. Yet participants rated the AI as less helpful and reported higher cognitive load while using the scaffold. The paper argues this tension reveals a productive friction that designers should embrace rather than eliminate.