A controlled study on Thai bar exam essay grading reveals that LLM judges do not neutrally reproduce human inter-rater disagreement — they converge overwhelmingly on the majority human interpretation. When a rubric ambiguity caused a genuine split among expert human examiners (2 vs. 1), 22 of 26 LLMs clustered with…
The Llamion project introduces KEPT (Efficient Knowledge Preservation for Transformation), a recipe for converting a non-Llama 14B model (Orion-14B) into the standardized Llama architecture while preserving capabilities with minimal retraining. The approach combines parameter-preserving mappings and a training-free…
Existing semantic-level watermarking (SWM) schemes treat sentences as atomic units, making them vulnerable to paragraph-level paraphrase attacks that scramble sentence order and disrupt watermark signals. SAMark addresses this by anchoring watermark detection in a sentence-order-independent "green region" in semantic…
TAPO (Trait-Aware Policy Optimization) is a post-training framework for autoregressive multi-trait automated essay scoring (AES) that decomposes reward signals along both sample and trait dimensions, rather than using a single scalar reward. It integrates four reward components — global scoring consistency,…
This paper proposes using LLMs as the semantic-matching engine for content-based publish/subscribe brokers in agentic AI systems spanning edge-cloud continua, overcoming the vocabulary and modality limitations of keyword and embedding filters. The authors identify two critical crossover thresholds: a context-window…
Double Triangle Annotation is a two-layer human-in-the-loop framework that exploits error independence between architecturally distinct Multimodal LLMs to auto-accept annotations where models agree, routing only conflicts to human reviewers. The design sidesteps LLM hallucination risks and avoids task-specific…
A study fine-tuning DeBERTa on a corrected multi-source PIIBench dataset spanning 82 entity types finds that direct token classification fine-tuning consistently outperforms more complex hierarchical and curriculum-based variants. On a 100K-record held-out evaluation, direct fine-tuning achieves F1 0.6455 versus…
Out-of-the-box language models exhibit significantly more reliable memory than real humans, undermining their validity as user simulators in applications like education or HCI. Wang et al. benchmark LLMs against humans on classic psychology memory tasks and find that even explicit prompting to behave humanly is…
Current LLM evaluation benchmarks for user profiling are limited to static data snapshots, failing to capture the dynamic, continuously evolving nature of real-world personalization systems. StreamProfileBench addresses this gap with a large-scale benchmark of 120,000+ UGC posts from 7,000+ users across five…
Across 10 LLMs, 3 benchmarks, and 68 experimental cells (~12,680 total inputs), meaning-bearing perturbations (paraphrase, synonym substitution) cause significantly more answer inconsistency in chain-of-thought and ReAct agents than surface-level formatting or reordering changes of equivalent severity — a gap…
Existing structured pruning methods fail on vision-language models (VLMs) because they are blind to chain-of-thought (CoT) reasoning dynamics and ignore activation distribution mismatches between visual and textual modalities. MuCRASP addresses this by targeting reasoning-critical components — specifically sparse…
The Thaka system won the KSAA-2026 Task 2 Arabic speech diacritization challenge by combining CATT-Whisper — a multimodal model pairing a character-level CATT text encoder with a frozen Whisper speech encoder — with a suite of regularization techniques designed to combat severe data scarcity (2,327 training samples,…
This paper empirically validates a mathematical creative quality metric ("Calibrated Surprise") at the engineering level using deliberately constrained conditions: a small base model and ~100 expert chain-of-thought annotations generated via the BC Protocol. The authors introduce Creative Quality Alignment (CQA) as a…
B³D-RWKV is a 7.2B-parameter model that resolves the fundamental architectural tension between causal linear models (unidirectional, O(L) inference) and discrete diffusion models (bidirectional attention requirement) through a novel triplet-block layout. By unifying RWKV's linear-time inference efficiency with…
SafeCtrl-RL introduces a reinforcement learning agent that dynamically selects prompt adjustment strategies at inference time to suppress unsafe LLM behavior, framing dialogue generation as a sequential decision process. Critically, it requires no model retraining or parameter modification, positioning it as a…
Persistent LLM agents that store memory as unstructured flat text suffer from "provenance-role collapse" — a failure mode where the agent loses track of the source and epistemic status of recalled information, leading to source-monitoring errors. MemIR (Memory Intermediate Representation) addresses this…
Legal LLMs systematically fail at temporal reasoning because they anchor to their training cutoff and don't constrain search queries by time — a critical flaw given that applicable law must match the temporal context of a case. LegalSearch-R1 addresses this with an end-to-end reinforcement learning framework combining…
Domain-adaptive continued pretraining (DAPT) on learner corpora does not uniformly improve transformer-based automated essay scoring (AES): full-corpus DAPT on EFCAMDAT yields mixed results across models and datasets, largely due to mismatches in proficiency level, genre, and communicative purpose. However, a…
TIAR (Trajectory-Informed Advantage Reweighting) extends ternary-reward-based abstention learning by dynamically reweighting the abstention advantage during GRPO training, using the distribution of sampled rollout trajectories as an implicit confidence signal over the model's knowledge boundaries. Rather than applying…
A new paper identifies a systematic dissociation in LLMs between internal causal representations and verbalized outputs, termed "Causal Tongue-Tie." On anti-commonsense causal reasoning items, a simple linear probe recovers the correct answer from hidden states with ~0.97 accuracy, while the model's Yes/No output…
A large-scale factorial grid search (1,345 total runs across two phases) finds that the optimal learning-rate warmdown fraction (33%) is invariant to bit-width across FP16, INT8, and INT6 quantisation-aware training for decoder language models in the 5M–350M parameter range, falsifying the hypothesis that…
This paper diagnoses failure modes in NLI-checker-guided reinforcement learning for medical RAG systems, showing that a checker's training-time output distribution is the critical variable — not its held-out accuracy. Using GRPO-trained agents (Qwen2.5-7B, Qwen3-4B, Llama-3.1-8B) across four medical QA benchmarks, the…
Merging monolingually pre-trained language models to achieve multilingual capability does not work — it causes performance collapse due to cross-language interference. Unlike fine-tuning, where model merging has shown flexibility and success, pre-training produces language-specific internal representations that are…
TTPrint introduces a two-phase architecture for extracting MITRE ATT&CK techniques from cyber threat intelligence (CTI) reports: a divergent phase that decomposes reports into atomic behaviors and broadly proposes candidate techniques, followed by a convergent verification phase that anchors candidates to source-text…
Reinforcement Learning with Verifiable Rewards (RLVR) requires ground-truth labels for reward computation, but labeling costs make full annotation impractical. Unsupervised RLVR workarounds using pseudo-labels suffer from training collapse. This paper introduces RLAVR, a framework that uses active learning to…
PolyGnosis 2.0 is a multi-agent system that fuses Polymarket anomaly signals with GDELT OSINT streams to identify "Perspective Mismatches" — narrative divergences between prediction market sentiment and global media — as high-alpha trading signals. The paper rigorously benchmarks agentic harness engineering techniques…
Current activation verbalization methods are siloed — each model can only explain its own internal representations. UAV breaks this constraint by training a lightweight adapter that projects activations from arbitrary "donor" models into the embedding space of a shared decoder, enabling cross-model, cross-family…
Current LLM-based agents are strictly reactive, computing only after explicit user prompts and wasting the idle time between interactions. ProAct is a proactive agent architecture that uses idle-time compute to analyze dialogue history and persistent memory, anticipate upcoming user needs, and pre-fetch relevant…
Inductive Deductive Synthesis (IDS) is a novel agentic LLM framework that jointly and incrementally co-synthesizes implementation and formal proof for distributed systems, learning from failed attempts to guide strategy selection. It solves all 7/7 distributed key-value-store specifications—compared to 2/7 for SOTA…
The Preisach Attention Layer (PAL) replaces softmax attention with a binary relay operator rooted in the classical Preisach hysteresis model, maintaining a stack of local extrema as internal state. A single-layer PAL-Transformer achieves Turing-completeness at O(1) depth — outperforming standard hard-attention…
Masked diffusion language models suffer from a fundamental inability to capture inter-token correlations, forcing a quality-throughput tradeoff. DiLaDiff addresses this by introducing a three-stage architecture: a semantically-rich continuous latent space (via fine-tuned auto-encoder), a latent diffusion prior, and a…
Open-world object detection (OWOD) systems are severely polluted at inference time: fewer than 10% of "unknown" predictions are genuine future-task objects, while 46–71% are background false positives. The root cause is an information bottleneck at the objectness head, not missing information — high-dimensional…
Aircraft disassembly at end-of-life is a large-scale combinatorial scheduling problem with industrial significance but thin profit margins, requiring formal optimization to be economically viable. Thomas and Schaus formulate the Aircraft Disassembly Scheduling Problem (ADSP) and benchmark two solving approaches —…
PCSP (Persona Conditioned Shared Policy) demonstrates that a single reinforcement learning policy, conditioned on frozen LLM embeddings of natural-language persona descriptions, can generate behaviorally distinct and consistent NPCs at scale. The architecture combines low-rank persona projection with a PPO + InfoNCE +…
This paper proposes a hybrid optimization framework that embeds Constraint Programming (CP) as a subroutine within a Dynamic Programming (DP) search, applied to the Partial Shop Scheduling Problem (PSSP). Rather than running CP and DP as competing paradigms, CP's global constraint propagation is used to prune the DP…
Any2Any is a transfer learning paradigm that adapts pretrained whole-body tracking (WBT) policies to new humanoid robot embodiments without training from scratch. It combines kinematic alignment — to reconcile input/output spaces between source and target robots — with parameter-efficient fine-tuning (PEFT) on…
Intrinsic multimodal knowledge editing in MLLMs reliably updates facts but consistently fails to generalize edits across semantically equivalent visual and linguistic variations — a problem the authors trace to missing semantic supervision, rigid edit scopes, and single-sample anchoring in high-dimensional spaces.…
High-resolution image perception is a core bottleneck for multimodal LLMs, where existing visual search methods force a tradeoff between full coverage (scan-based, computationally expensive) and efficiency (expert-assisted, prone to blind spots). CVSearch resolves this with a training-free "Assess-then-Search"…
MetaEvaluator is a model-agnostic framework that uses meta-learning over a pool of reference models to produce a transferable initialization, enabling accurate performance estimation of new, unseen models on entirely unlabeled datasets. It addresses a critical bottleneck in the ML ecosystem: the cost and scalability…
HARNESS-LM (HLM) is a three-phase training framework that distills a billion-parameter SLM retriever into a sub-600M (deployed at 190M) parameter student model for sponsored search via: (1) fine-tuning a large SLM as a teacher, (2) L2-based query representation alignment for knowledge transfer, and (3) contrastive…
A longitudinal study spanning over a decade of Android applications demonstrates that temporal distribution shift (concept drift) meaningfully erodes the adversarial robustness of malware detection models — not just their clean accuracy. Three deployment protocols were evaluated across multiple classifier families…
Reinforcement learning agents trained sequentially on multiple tasks exhibit structured, predictable out-of-distribution generalization behavior that is not random but shaped by training history. Brown & Young introduce *latent policy gradients*, a method that models the evolution of low-dimensional latent variables…
Co-ReAct addresses a core weakness of ReAct-style agents — reliance on internal judgment for action selection — by injecting dynamically generated rubrics at each decision step during inference, not just as post-hoc evaluators. A dedicated rubric generator is trained with GRPO using a listwise Spearman…
EM-Vid addresses a core failure mode in autoregressive multi-shot video generation: full-frame memory reuse conflates persistent entity appearance with transient scene context, causing information leakage and high compute costs. The proposed system replaces full-frame storage with an entity-indexed bank of latent…
Current LLM conversational systems are purely reactive and face a hard efficiency–quality tradeoff when handling multi-turn dialogue history: full-history concatenation scales token cost linearly, while truncation destroys cross-turn context. OnePred sidesteps this by maintaining a recursively updated, compressed…
PhotoFlow introduces a three-role agent architecture (Director-Reviewer-Reflector) that enables closed-loop camera search in arbitrary 3D Blender scenes given only a language intent — no pre-selected pose or reference image. The system addresses the dual challenge of 3D spatial reasoning and aesthetic judgment, areas…
LLM agents with persistent memory are vulnerable to adversarial poisoning via ordinary user interactions — a threat class that existing online defenses (prompt filtering, output blocking) fail to address retroactively. MemAudit introduces a post-hoc auditing framework that combines counterfactual causal scoring of…
Claude Code deployed in an agentic proving loop achieves near-ceiling performance on CLEVER, a Lean 4 benchmark for verifiable code generation, exposing a growing capability-benchmark mismatch in formal program verification. The study demonstrates that tight compiler-in-the-loop feedback enables LLMs to not only…
Subliminal learning — the transfer of task-relevant knowledge from teacher to student models via distillation on task-unrelated inputs — has previously been attributed to closely matched initializations. This paper refutes that assumption, demonstrating instead that compatible output heads (specifically auxiliary…
Ganguli & Constantinescu propose a structured Bayesian prior built on a disentangled deep generative model whose latent space is explicitly partitioned into interpretable physical parameters and residual variability. By linearizing the generator, they derive conditions under which the posterior achieves approximate…
New compiled intelligence delivered to your inbox.