Speculative decoding draft models face severe KV cache bottlenecks during long-output generation, but standard prefill-based compression methods lose relevance as decoding progresses. This paper identifies that "decode staleness" phenomenon and proposes RIFT, a family of selection-based KV compression strategies that balance important prefill tokens with a recent-token sliding window, plus adaptive decode-time refresh and a complementary compression-aware distillation training phase. The core claim is that these methods allow small external drafters to retain over 96% of their full-cache acceptance length under tight memory budgets.

This paper tackles certified neural approximation of the Dirichlet Monge--Amp\`ere equation, a fully nonlinear PDE whose solutions must be convex. The authors study implicit-depth equilibrium networks built from softplus units with nonnegative weights, proving that a resolvent--curvature cancellation improves the Hessian growth bound from $(1-q)^{-3}$ to the sharp exponent $(1-q)^{-2}$. They then derive residual-based a posteriori estimates and validate them with interval-arithmetic certificates on trained models, yielding rigorous uniform evaluation guarantees for implicit neural PDE solvers.

Fine-grained image-text retrieval requires aligning local semantics such as objects, attributes, and relations, yet fixed patches and tokens often produce fragmented matching units. This paper proposes HFCMA, a framework that dynamically aggregates low-level patches and tokens into adaptive visual regions and textual phrases, and then aligns these sets with entropy-regularized optimal transport under a frozen CLIP dual-encoder architecture. By making semantic unit construction an explicit learnable step and enforcing globally consistent matching via a transport plan, HFCMA aims to improve discrimination of subtle differences without sacrificing retrieval scalability.

Guardrails for LLMs typically fail when the same safety case appears across different interfaces—prompts, responses, agent traces, or policy-conditioned inputs. UniGuard reframes guardrail training as multi-view representation learning, where prompts, conversations, and agent trajectories are treated as semantically aligned views of an underlying safety case. A compact 0.6B model is trained with supervised classification, contrastive alignment, and asymmetric confidence calibration, showing large F1 and calibration gains on unseen-interface benchmarks while remaining competitive with proprietary models.

DuoCook evaluates whether a single vision-language model can centrally coordinate two cooking agents in Unreal Engine kitchens inspired by Overcooked. The benchmark spans 20 maps across five categories—basic, shared equipment, dynamic terrain, combined dynamic/shared, and separated workspaces—requiring the model to assign high-level actions from top-down visual observations while the engine handles low-level navigation. A memory ablation tests the value of a shared, self-authored belief notebook. Performance is far from saturated even with heavy scaffolding, suggesting current VLMs remain poor at real-time centralized coordination, though the task interface provides substantial structured assistance.

Language models fail to learn shortest-path search on unseen symbolic graphs as graphs scale beyond about 50 nodes, a wall this paper diagnoses and breaks through. The authors show that next-token prediction provides too little and too skewed supervision, while standard tokenization introduces routing and permutation artifacts. By combining multi-token prediction—which densely supervises all sub-paths and implicitly mixes hop lengths—with structural annotation (STAN) and a node list (+NN), they stably train small transformers on graphs up to 125 nodes with far lower variance. The work makes a strong case that objective design and input formatting matter more than model size for latent multi-hop reasoning.

The paper studies how much embedding work is wasted when documents in a RAG knowledge base are revised, comparing fixed-size, recursive paragraph-first, and content-defined chunking on 120 adjacent Wikipedia revision pairs in Chinese and English. It finds that fixed-size windows force the most re-embedding (57–68\% of the embedding-input workload), while content-aware methods cut this by double-digit percentage points. Surprisingly, the extra recomputation buys almost no retrieval recall advantage under a fixed 2048-token budget. The authors argue that update cost, measured under byte-identical reuse, should be reported alongside static retrieval metrics when choosing a chunker for an evolving knowledge base.

Hierarchical language models split semantic planning from token generation, but prior work leaves the latent space and decoder mismatched. Cola2 co-trains a dynamic compression VAE and a masked-diffusion decoder so that token-prediction gradients shape the encoder while a separate DiT fits the latent prior via stop-gradient. The result is a compact continuous representation that stays aligned with the actual text during block-wise inference, and scaling experiments up to about 210B tokens show competitive, unsaturated curves. A 346B-source-token checkpoint reaches a 60.10 eight-task average, though the model remains far from converged.

This survey argues that the central bottleneck for advancing remote sensing vision-language models has shifted from architectural innovation to data curation. It introduces GeoChef, an end-to-end framework that maps dataset taxonomy, construction pipelines, fine-tuning recipes, and evaluation protocols onto a culinary metaphor. By cataloging over 120 datasets and launching a community-driven warehouse, the authors aim to turn small academic teams into proactive data chefs rather than passive consumers of foundation models.

LLM agents with persistent external memory often leave derived artifacts—such as summaries—intact after a user deletes a source record, creating a gap between storage-level deletion and behavioral privacy. AfterDeleteBench introduces a paired behavioral audit that compares an exposed-then-deleted agent against a never-exposed twin to attribute post-deletion secret recovery to the original exposure. The paper shows that exact reconstruction can persist even when no verbatim copy of the secret survives, and that the response interface itself is part of the measurement system, making the benchmark a useful narrow probe into a broader agent-memory privacy problem.

Collective ordering in correlated electron systems produces emergent, non-Markovian dynamics that cannot be closed at the level of coarse-grained observables alone. This paper proposes a graph-recurrent neural network (GRNN) that fuses gated recurrent memory with graph message passing to learn finite-time evolution maps directly from trajectories, bypassing repeated microscopic force evaluations. Site-resolved hidden states encode unresolved history and communicate across neighboring sites via graph-coupled hidden messages, giving a scalable, translationally invariant update rule. The authors demonstrate the approach on charge-density-wave domain coarsening in the semiclassical Holstein model, showing quantitative agreement with kernel-polynomial-method reference dynamics.

Existing cultural bias benchmarks rely on costly manual annotation by native speakers, which limits multilingual scalability and introduces subjectivity. WikiOriginQA tackles this by using Wikipedia article creation dates across 342 language editions as an objective proxy for knowledge origin, then automatically generating short-answer QA pairs with LLMs. The resulting dataset supports the claim that LLMs are skewed toward Anglophone knowledge and that propagation into English Wikipedia is a key predictor of model performance.

The paper challenges the widespread assumption that reinforcement learning (RL) invents the reflection marker "wait" in reasoning models. Using historical-linguistics-inspired "grammaticalization" tracking, the authors trace the word through 120 open-model checkpoints and find it almost always arrives from training data—mid-training, fine-tuning on reasoning traces, or reasoning-rich pretraining—while zero-RL alone rarely coins it.

Extraction pipelines often ask language models to return JSON matching a schema that sets a minimum list length. This paper shows that when real documents contain fewer items than that minimum, open-weight models pad their lists---mostly with real text from the document that evades simple source checks. The authors preregister a design with 558 documents, eight GPU models, and paired contrasts, finding that an unenforced minimum raises over-counting from 34.5% to 55.4% and that common fixes fail or trade recall for precision. The work matters because schema-driven extraction is ubiquitous, and the padding is hard to catch without ground-truth answer lists.

Time-series dataset similarity is task-dependent: no single metric reliably predicts downstream transfer error across models and protocols. This paper proposes DISC, a learnable aggregation framework that calibrates a base mixture of standardized similarity metrics with a rank-based prior gate. On a shared 46-dataset split, DISC raises the average source-wise Pearson correlation with forecasting error by about 10% over the best single metric, supporting the claim that learned integration is a stronger proxy for transfer behavior than manual metric selection.

On-policy self-distillation improves reasoning by training a model on its own trajectories with dense token-level guidance from a reference-conditioned teacher. The authors identify a critical flaw: because any single reference is just one valid realization of reasoning, conditioning the teacher on the full reference makes the target distribution overly sensitive to that specific solution's wording, ordering, and path. They propose Monte Carlo Privileged Distillation (MCPD), which averages the teacher's next-token distributions over multiple randomly masked partial views of the same reference to produce a more stable marginal target. Across math and science benchmarks, MCPD consistently outperforms standard full-reference distillation.

This paper asks whether robot learning must rely on embodied data by training in-context action predictors entirely on procedurally generated, non-robotic mathematical abstractions—tabular MDPs, contextual bandits, and linear dynamical systems. Using a Prior-Data Fitted Network framework, the authors show that these synthetic control priors achieve non-trivial zero-shot action prediction on held-out simulated and real-world robots, including manipulators, wheeled vehicles, quadrupeds, and aerial platforms. The work matters because it cheaply decouples structural prior acquisition from expensive physical data collection, offering a scalable complement to embodied foundation model pipelines.

This paper proposes replacing the physics-based BSIM-CMG compact model with tiny MLPs (8–16 neurons) to predict drain current and charges for a 3 nm nanosheet, then embedding them in Verilog-A to perform a full standard cell library characterization and subsequent static timing analysis on industrial-scale circuits. The authors claim sub-percent delay and power errors alongside up to 8× faster library characterization, positioning the approach as a viable alternative to equation-based transistor models in modern EDA flows.

Large language models rely on just-in-time compilers like Triton to generate optimized GPU kernels, caching the results for reuse. This paper asks whether a Rowhammer-induced bit-flip can silently degrade inference efficiency without harming output correctness, and answers affirmatively with LLVMhammer. By corrupting a target-independent optimization pass in Triton, the attack forces emission of a suboptimal kernel that raises latency by 4–16%, while the on-disk cache makes the slowdown persist across restarts.

Learning discrete distributions from batched user data under person-level privacy is a fundamental problem in federated analytics. This paper develops the first polynomial-time pure-DP algorithm that matches the information-theoretic sample complexity up to logarithmic factors, via a new degree-6 sum-of-squares certificate based on a trace-bounded diagonal majorizer for multinomial covariance. It also proves a statistical-query lower bound suggesting the gap between efficient and optimal robust rates may be inherent.