This paper presents a large-scale comparative study of memorization across six open LLM families (Pythia, OLMo1/2/3, OpenLLaMA, StarCoder) ranging from 1B to 32B parameters. By analyzing both statistical patterns and internal mechanisms (attention heads, layer decoding), it identifies universal behaviors—such as log-linear scaling of memorization rates with model size and high compressibility of memorized sequences—while revealing family-specific signatures in memorization structure. The work bridges isolated findings from single-model studies to establish general principles of how transformers memorize training data.

EvoIdeator addresses the challenge of iteratively refining scientific research ideas using LLMs by bridging the gap between scalar RL rewards and coarse language feedback. The core innovation is a dual-signal approach combining lexicographic rewards with checklist-grounded, span-level language feedback integrated directly into the RL training loop using Dr. GRPO. This allows a 4B parameter model to outperform larger frontier models like Gemini 3 Flash and DeepSeek-V3.2 on scientific rigor criteria.

Medical text summarization helps clinicians process millions of biomedical articles, but fine-tuning large language models demands prohibitive resources. This paper compares Low-Rank Adaptation (LoRA), Prompt Tuning, and full fine-tuning across Flan-T5-Small, Base, and Large on PubMed summarization. The counter-intuitive finding is that updating fewer than 1% of parameters via LoRA consistently outperforms full fine-tuning, suggesting that low-rank constraints provide effective regularization.

Time toxicity—the cumulative healthcare contact days imposed by clinical trial participation—is an important patient-centric metric buried in dense Schedule of Assessments (SoA) tables. This work proposes TimeTox, a Gemini-based LLM pipeline that extracts time toxicity from protocol PDFs at scale, comparing a single-pass architecture against a two-stage structure-then-count approach. The authors deploy their system on 644 real-world oncology protocols and find that synthetic benchmark accuracy is a poor predictor of real-world reliability, a lesson critical for clinical NLP deployment.

This paper introduces SemEval-2026 Task 12, Abductive Event Reasoning (AER), a shared task requiring systems to identify the most plausible direct cause of a target event from noisy multi-document evidence. The task is cast as an evidence-grounded multiple-choice benchmark with multiple correct answers allowed, capturing challenges like distributed evidence, indirect background factors, and semantically related distractors. With 122 participants and 518 submissions, it represents a significant community effort to benchmark real-world causal reasoning in long-context settings.

ROM tackles overthinking in Large Reasoning Models, where models generate redundant reasoning after reaching correct answers. The core idea is a lightweight streaming detector—an 8.13M parameter head attached to late-layer hidden states of a frozen LLM—that predicts overthinking probability token-by-token and triggers early stopping. It matters because it promises 47% token reduction without full model retraining. We find the method empirically effective but note concerns regarding data scaling limits and labeling costs.

Dyadic is a web-based platform for studying human-human and human-AI conversations through text or voice-based interaction. It attempts to solve the methodological gap in conversation research by providing turnkey tools for experimental manipulation, live monitoring, and in-situ survey delivery during ongoing chats. The core value proposition is lowering barriers to entry for researchers studying dyadic interaction processes without requiring programming expertise.

Cross-Layer Transcoders (CLTs) compress the attribution graphs used in mechanistic interpretability by sharing features across transformer layers, but their quadratic parameter scaling ($N_{\text{CLT}} \propto L^2$) makes training and analysis prohibitively expensive for most researchers. This paper introduces CLT-Forge, an open-source library that combines feature-sharded distributed training, compressed activation caching (int8/int4/int2 with zstd), automated interpretability pipelines, and integration with Circuit-Tracer to provide the first unified workflow for end-to-end CLT analysis at scale.

WorldCache addresses the prohibitive latency of Diffusion Transformers (DiTs) for video world models by replacing static feature caching with a content-aware dynamical approximation framework. The method introduces motion-adaptive thresholds, saliency-weighted drift estimation, and optimal feature blending to eliminate ghosting artifacts during fast motion. Achieving 2.3× speedup on Cosmos-Predict2.5 with 99.4% quality retention, it offers a training-free path toward interactive world simulation.

Large language models often lack coverage in specialized, data-scarce domains where web text is limited. This paper proposes SPA (Scaling Prompt-engineered Augmentation), a baseline that generates large-scale synthetic corpora using just seven carefully designed prompt templates grounded in cognitive learning strategies (Concept Learning, Critical Thinking, and Generative Learning). The core finding is that this simple approach consistently outperforms complex RL-based methods like SEAL and multi-stage pipelines like EntiGraph across Wikipedia QA, long-document comprehension, and multi-hop reasoning benchmarks, suggesting that careful prompt design combined with straightforward scaling is surprisingly effective for knowledge injection.

Standard Transformers apply fixed-depth computation regardless of problem difficulty, limiting their ability to solve tasks requiring variable-depth reasoning like multi-hop traversal or nested logic. This paper proposes a depth-recurrent Transformer that iteratively applies a shared-weight block in latent space—enabling 'vertical Chain-of-Thought' where models trade recurrence steps for deeper reasoning without consuming context window. The work demonstrates strong compositional generalization on three synthetic tasks and offers a mechanistic alternative to horizontal token-generation paradigms.

Long-context LLM inference hits a memory wall: each decode step requires scanning the entire KV cache, incurring $O(n)$ memory bandwidth that cannot be solved by faster arithmetic. PRISM proposes a thin-film lithium niobate photonic accelerator that performs the block-selection similarity search in $O(1)$ optical latency using a broadcast-and-weight architecture, eliminating the $O(n)$ scan entirely. The work claims $16\times$–$32\times$ traffic reduction at 64K–128K tokens and a four-order-of-magnitude energy advantage over GPU baselines by matching photonic hardware capabilities—passive query broadcast, quasi-static microring weights, and low-precision rank output—to the selection task.

This paper examines representation genesis—the transition from non-representational physical systems to those with content-manipulable states. It argues that major frameworks in philosophy of mind (Language of Thought, teleosemantics, predictive processing, enactivism, and genetic phenomenology) share a 'Representation Presupposition' structure that prevents them from explaining this first acquisition without circularity. With large language models now achieving high cognitive performance without clear genesis events, the absence of a satisfactory theory becomes urgent.

The paper critiques the institutionalization of LLM benchmarks as "Silicon Bureaucracy" and "AI Test-Oriented Education", arguing high scores often conflate exam-oriented competence with genuine generalization due to data contamination. It proposes an audit framework using a router-worker setup: clean-control routers transmit full questions while noisy routers delete, rewrite, and perturb before aggregation. For clean benchmarks, noisy aggregation should not systematically exceed the baseline; persistent above-baseline gains suggest contamination-related memory activation. The core finding—that 10 of 12 models exceed clean baselines under multi-router noisy conditions—challenges the interpretability of raw benchmark scores.

CAID tackles long-horizon software engineering tasks where single agents struggle with accuracy and wall-clock time. The core idea is Centralized Asynchronous Isolated Delegation: a manager decomposes tasks into dependency graphs and delegates to multiple engineer agents working in isolated git worktrees, integrating progress via branch-and-merge. The system improves accuracy by 26.7% absolute on PaperBench and 14.3% on Commit0, demonstrating that structured coordination grounded in SWE primitives outperforms simply scaling single-agent iteration budgets.

AdaRubric solves the static-rubric bottleneck in LLM-as-Judge evaluation by dynamically generating task-specific evaluation dimensions from task descriptions. It scores agent trajectories step-by-step with confidence-weighted per-dimension feedback and filters preference pairs using the DimensionAwareFilter—a provably necessary mechanism to prevent high-scoring dimensions from masking failures. The approach achieves Pearson $r=0.79$ correlation with human judgments and yields substantial downstream gains: +6.8–8.5 percentage points in DPO task success and +6.6 pp faster PPO convergence at 5K steps.

This paper addresses computational barriers for Brazilian Portuguese question answering by systematically evaluating Parameter-Efficient Fine-Tuning (PEFT) methods on BERTimbau models using the SQuAD-BR dataset. The authors test LoRA, DoRA, QLoRA, and QDoRA across Base (110M) and Large (335M) variants, demonstrating that LoRA achieves 95.8% of full fine-tuning performance while reducing training time by 73.5%. A key finding is that PEFT methods require substantially higher learning rates ($2\times 10^{-4}$) than standard BERT fine-tuning to achieve optimal results, with quantization resilience favoring larger models.

This paper tackles the problem of measuring dialectal bias in LLMs for Bengali, a low-resource language with nine major regional variants. The authors propose a two-phase framework combining RAG-based translation to create dialectal benchmarks with an RLAIF-inspired evaluation protocol that uses CoT-first reasoning and multi-judge validation. They expose the catastrophic failure of traditional metrics like BLEU and WER for agglutinative dialectal Bengali, showing that LLM-as-judge better predicts human quality assessments.

Large Language Models often inherit societal biases that manifest as stereotyped associations across demographic groups. This paper proposes CatRAG, a dual-mechanism debiasing framework that combines a category-theoretic functor-guided projection—collapsing protected-attribute directions in embedding space via spectral decomposition—with diversity-aware Retrieval-Augmented Generation to ground inference in balanced evidence. Evaluated on the BBQ benchmark across Llama-3, GPT-OSS, and Gemma-3, the method claims to reduce bias scores from ~60% to near zero while improving accuracy by up to 40% over base models.

AgentHER tackles the data waste problem in LLM agent training by adapting Hindsight Experience Replay (HER) from RL to natural-language trajectories. The core insight is that failed trajectories—typically 60–75% of collected data—often represent valid demonstrations for achievable alternative goals. The paper proposes a four-stage pipeline with multi-judge verification that converts discarded failures into SFT and DPO training data, yielding +7.1–11.7 pp gains over success-only fine-tuning across four model families on WebArena and ToolBench.