AI Research

Scaling Laws and Fundamental Paradigms

A central methodological debate persists regarding paths to AGI. Ilya Sutskever posits that while continued scaling yields improvements, it will ultimately "stall hard," necessitating a new paradigm to overcome the "biggest AGI blocker": models generalize approximately 100 times worse than humans [1]. This challenges the scaling orthodoxy documented in OLMo research, which suggests smaller models fail on rare tasks due to gradient interference—where high-frequency tasks monopolize neurons—while larger models escape this through reduced interference [54]. Critics note this mechanism may conflate parameter count with architectural geometry [C54].

OpenAI's Greg Brockman reveals that the scaling hypothesis was discovered empirically during the Dota 2 project, where doubling compute consistently doubled performance without plateauing [3]. This observation led OpenAI to build its strategy "backwards," releasing a general API and following emergent use cases [3]. Brockman predicts AI will solve at least one Millennium Prize Problem within two to five years [3]. However, critics argue this retrospective narrative suffers from survivorship bias and post-hoc rationalization, noting that the scaling hypothesis had theoretical and empirical precursors well before OpenAI's Dota work [Counter-claims]. The prediction of solving Millennium Prize Problems is seen as an example of tech-industry overconfidence without rigorous evidentiary basis, ignoring the need for novel conceptual breakthroughs [Counter-claims].

Research on the "second scaling law" indicates that increasing token limits for reasoning models yields substantial gains—expanding Codex limits from 3M to 10M tripled autonomous work capacity [79]. However, critics counter that such gains may plateau on tasks requiring complex judgment [C79]. Ethan Mollick notes that capability growth now includes test-time compute and parallelism, providing multiple paths to improvement [4], though critics point to revised Chinchilla laws and diminishing returns [Counter-claims].

New research introduces InfoLaw, a data-aware scaling framework that predicts loss from consumed tokens, model size, data mixture weights, and repetition [1]. It models pretraining as information accumulation, where quality controls information density and repetition induces scale-dependent diminishing returns. InfoLaw predicts performance on unseen data recipes and larger scale runs (up to 7B, 425B tokens) with 0.15% mean and 0.96% max absolute error in loss, and extrapolates reliably across overtraining levels [1].

Reasoning Efficiency and the Epiphenomenality of Chain-of-Thought

Recent work challenges the assumption that longer chain-of-thought (CoT) traces correspond to deeper computation. Research identifies a "commitment boundary" in large reasoning models—a sharp, often single-step transition from transient intermediate guesses to a stable, high-confidence answer that occurs well before the reasoning block ends [New-39]. After this boundary, subsequent CoT steps appear epiphenomenal: they leave the final answer probability unaltered, suggesting that much of extended reasoning may be post-hoc rationalization rather than causally necessary computation [New-39]. Attention probes can linearly decode these answer-formation stages from intermediate representations with high accuracy, generalizing to unseen tasks [New-39]. Exploiting this signal for early-exit strategies reduces CoT length by up to 55% on average without meaningful performance loss [New-39]. However, skeptics caution that defining the boundary as a single step may oversimplify a potentially gradual process, and that steps after commitment could still reinforce answer robustness or certainty in ways not captured by final-answer probability alone [Counter-claims].

Base Large Language Models (LLMs) from 2023-2024 are increasingly recognized as lacking fluid intelligence and mathematical reasoning capabilities, a fact now widely accepted despite initial controversy [1]. This contrasts with emerging Language Reasoning Models (LRMs), which are hypothesized to perform significantly better on complex reasoning tasks [1]. Critics argue that advocates for early LLMs failed to recognize these limitations due to preconceived notions [1]. However, counter-arguments suggest that this claim overgeneralizes from limited evidence, that LRMs' capabilities are speculative without direct testing, and that attributing oversight solely to preconceived notions ignores legitimate evidence of emergent reasoning in early LLMs [Counter-claims].

Architectural Innovations and Inference Efficiency

Neuroscience continues to inspire architectures, with Hubel-Wiesel modules modeling perception-memory interfaces [170]. Yann LeCun asserts that the Joint Embedding Predictive Architecture (JEPA) remains the primary architectural goal for AI development, contrasting it with the Causal World Model (CWM) which he characterizes as a token-generative baseline [5]. Critics argue that declaring JEPA as the singular goal is premature, given the field's paradigm shifts and alternative architectures, and that labeling CWM as 'just a token-generative baseline' undersells its contributions to causal reasoning [Counter-claims].

Training efficiency advances include Piper, which achieves 2–3.5× higher Model FLOP Utilization via optimized pipeline schedules [93], though critics suggest the range reflects cherry-picked configurations [C93]. EMO introduces a 1B-active, 14B-total MoE model enabling retention of 25% of experts with 1% performance drop [118]. Google's TurboQuant combines PolarQuant with Quantized Johnson-Lindenstrauss for 6× KV cache reduction [53], operating as a software-only solution without retraining [6].

State Space Models (SSMs) represent a prominent alternative to transformers. Mamba-3 shifts focus to inference efficiency, demonstrating faster prefill and decode latency than transformers at 1.5B scale [148]. Critics argue latency comparisons may be hardware-specific [Counter-claims].

Speculative decoding has emerged as critical for inference acceleration. FailFast leverages Diffusion LLMs as drafters, achieving up to 4.9× speedup [147]. However, skeptics note "up to" speedups represent cherry-picked scenarios [Counter-claims]. SpecBlock proposes a block-iterative drafter improving mean speedup by 8-13% over EAGLE-3 at 44-52% of its drafting cost [New-2].

Theoretical analysis of KV cache compression in Transformers reveals depth-cache tradeoffs for k-hop pointer chasing on n tokens with cache size s. A conjectured product lower bound L = Ω(⌈k/s⌉ · ⌈log₂n/(Hmp)⌉) is proposed for locality-respecting Transformers [6]. This is supported by matching upper bounds achieved by windowed pointer doubling [6]. However, the product bound is conjectural, with only a weaker max-bound proven unconditionally, and the upper bound may not strictly respect locality constraints [Counter-claims]. Per-window distinguishability counting lower bounds cannot exceed ⌈k/s⌉ once Hmp ≥ log₂n, indicating a bandwidth barrier [6]. Adaptive locality-respecting caches achieve an exact Pr[ℰ] = s/n error, explaining the dominance of heavy-hitter eviction [6].

Token Pruning for Efficiency

Token pruning has emerged as a strategy to boost efficiency in multilingual LLMs, particularly for domain-specific applications. By eliminating tokens and embeddings for irrelevant languages, it aims to improve generation stability by reducing language confusion and can enhance performance on specific tasks like Korean machine translation [2]. This method also reduces vocabulary size, making it suitable for memory-constrained deployments [2]. However, the impact on instruction-following capabilities can vary depending on the model architecture, suggesting a complex interaction with latent cross-lingual representations [2]. While token pruning offers modest gains in inference latency, critics argue that the benefits might be overstated, and the method could introduce new challenges if critical domain-specific tokens are inadvertently removed [Counter-claims].

Visual Encoding and Geometric Fragility in Vision-Language Models

Despite strong semantic understanding, state-of-the-art Vision-Language Models (VLMs) exhibit systematic failures in geometric robustness, lacking spatial invariance and equivariance under simple rotations, scaling, and identity transformations [243]. Performance drops sharply as semantic content becomes sparse, observed across architectures and model capacities [243]. This reveals a systematic gap between semantic understanding and spatial reasoning, highlighting the need for stronger geometric grounding in multimodal systems [243].

Recent work on visual encoding in Multimodal Large Language Models (MLLMs) challenges the prevailing practice of global encoding followed by post-ViT compression. Controlled experiments show that slice-based encoding, which preserves local details through sliced views, outperforms global encoding across benchmarks. Additionally, intra-ViT early compression, reducing tokens in shallow ViT layers, substantially lowers visual-encoding FLOPs while maintaining performance. LLaVA-UHD v4, integrating these approaches, reduces visual-encoding FLOPs by 55.8% while matching or surpassing baselines on diverse tasks including document understanding, OCR, and general VQA [5].

Reinforcement Learning and Control

Fundamental Unlearnability in RLVR

Recent research reveals fundamental unlearnability in Reinforcement Learning with Verifiable Rewards (RLVR): among hard examples that models initially fail, a substantial subset remains unlearnable even with correct rollouts present, due to low gradient similarity and ungeneralizable reasoning patterns [239]. Cross-example gradient analysis shows that data augmentation does not improve gradient similarity, suggesting representation flaws are difficult to mitigate in RL [239]. This indicates fundamental limitations in current RL approaches for reasoning tasks.

Safety Alignment and the Alignment Tax

Safety post-training can improve policy compliance but often reduces general utility, described as the alignment tax. Orthogonal Gradient Projection for Safety Alignment (OGPSA) mitigates this by estimating a low-rank reference subspace from general-capability data and removing from safety gradients components lying in this subspace [238]. The resulting update is the steepest local safety-descent direction subject to first-order preservation constraints on reference objectives [238]. Across SFT and sequential SFT→DPO settings, OGPSA improves the safety-utility trade-off, with average performance gains increasing from 33.98% to 42.74% on Qwen2.5-7B-Instruct [238].

Multi-Response Reward Modeling

Conventional discriminative reward models evaluate each response independently, requiring multiple forward passes. Recent work introduces a multi-response design that concatenates responses with separator tokens and applies cross-entropy over their scalar scores, enabling direct comparative reasoning and efficient N-way preference learning with up to N× wall-clock speedup [240]. Built on a 4B vision-language backbone, this approach achieves state-of-the-art results on multimodal reward benchmarks and improves training stability when used with GRPO [240].

Inference-Time Evolution vs. Distillation

Evoflux demonstrates that inference-time evolutionary search over executable tool workflows outperforms supervised fine-tuning (SFT) and DPO for compact language model agents when teacher trace budgets are scarce [248]. Rather than distilling recovery behavior from few hundred traces—which rarely covers the repair space needed for failed plans over changing tool catalogs—Evoflux evolves typed workflow graphs through structured edits, execution feedback, and diversity pruning [248]. On MCP-Bench with 250 tools, it raises execution feasibility from ~3% to 17-24% for small planners, while SFT and DPO match, underperform, or collapse below zero-shot performance [248]. However, critics note that the comparison assumes scarce trace budgets, and distillation methods might perform comparably with larger budgets or different hyperparameter tuning [Counter-claims].

Edge-Native Reasoning for Decentralized Consensus

Research on edge-native Small Language Models (SLMs) for Decentralized Autonomous Organizations (DAOs) reveals a severe compute-accuracy inversion when applying System 2 reasoning [246]. On Sentinel-Bench, an autoregressive System 1 baseline achieved 100% adversarial robustness and state finality in under 13 seconds, while System 2 reasoning introduced a 26.7% Reasoning Non-Convergence rate, degrading consensus stability to 72.6% and imposing 17× latency overhead [246]. The study concludes that for edge-native SLMs under Byzantine Fault Tolerance constraints, System 1 parameterized intuition is structurally superior to System 2 iterative deliberation [246].

Phase-Wise Hallucination Mitigation

Large Vision-Language Models struggle with vision hallucination, where responses are inconsistent with visual input. Phase-wise Self-Reward Decoding (PSRD) enables dynamic hallucination mitigation at inference time without external supervision [245]. Visual hallucination exhibits phase-wise dynamic patterns, peaking at the onset of each semantic phase [245]. PSRD uses a distilled lightweight reward model to provide on-the-fly guidance, significantly reducing hallucination rates by 50.0% on LLaVA-1.5-7B across five benchmarks [245].

State Value Estimation in LLM RL

Accurate state value estimation, critical for stable RL training, remains an underexplored challenge in LLM post-training. Critics in standard LLM RL frameworks like PPO often degenerate into coarse group-average baselines rather than producing meaningful per-state estimates [10]. To address this, two new techniques have been proposed: Numca, which uses numerical spans as gradable milestones for value estimation, and Hista, which leverages LLM hidden states to compute weighted averages over disjoint rollouts and their returns [10]. Both methods improve value estimate accuracy and downstream training performance across multiple RL algorithms and model sizes with minimal added computational cost [10]. However, the claim that this area is 'underexplored' is debated, given existing work on token-level credit assignment, and the domain-specificity of Numca's numerical span approach may limit its generalizability [Counter-claims]. The computational overhead of Hista's buffer management is also questioned [Counter-claims].

Momentum-Based Reward for Traffic Signal RL

A Momentum-Based Reward Function (MBRF) for deep reinforcement learning (DRL) agents controlling urban traffic signals has been introduced, shifting the optimization objective from penalizing congestion to incentivizing continuous vehicle movement [9]. Evaluated in the SUMO traffic simulator, MBRF achieves better throughput-emission trade-offs and more stable training dynamics compared to conventional delay- and queue-based reward formulations, as well as classical non-learned controllers (Max Pressure, LQF) [9]. This suggests reward design is a critical lever for policy stability and environmental performance in adaptive traffic signal control [9]. Critics, however, question the generalizability of these simulation results, noting that momentum-based rewards might inadvertently worsen side-street emissions and that the outperformance of classical controllers may be due to specific tuning rather than intrinsic superiority [Counter-claims]. The claim that existing delay- and queue-based rewards are 'short-sighted or unstable' is also challenged as an overgeneralization [Counter-claims].

Probabilistic Safety Guarantees for Perception-Uncertain Agents

New methods are emerging to provide probabilistic safety guarantees for autonomous systems operating under imperfect perception. One approach models perception uncertainty as an Interval POMDP, using confidence intervals derived from finite labeled data on sensor misclassification rates [6]. This allows for the computation of a conservative belief set over states, which is then used to construct a runtime shield that blocks unsafe actions [6]. This shield offers finite-horizon safety guarantees: with high probability over training data, if true perception uncertainties lie within learned intervals, admitted actions meet a specified safety lower bound [6]. Experiments across four case studies demonstrate superior safety over state-of-the-art baselines [6]. However, critics caution that such guarantees are heavily reliant on the assumption that true perception rates remain within the learned intervals, which may not hold under real-world distribution shifts, and that the 'conservative' belief set might lead to overly cautious and inefficient behavior [Counter-claims].

LEPO (Latent Exploration Policy Optimization) is a novel framework that injects controllable stochasticity into the latent reasoning of LLMs using Gumbel-Softmax, preventing deterministic collapse and restoring exploratory capacity [8]. This allows RL to be applied directly to continuous latent representations, maintaining stochastic trajectory sampling during rollouts and unified gradient estimation for latents and tokens during optimization [8]. Experiments show LEPO significantly outperforms existing RL methods for both discrete and latent reasoning [8].

Agentic Systems and Autonomous Capabilities

April-May 2026 saw significant frontier releases. OpenAI's GPT-5.3-Codex achieved state-of-the-art performance on SWEBench Pro with 25% faster inference [7]. OpenAI has also released GPT-5.5, optimized for complex goals, tool use, self-verification, and multi-step task completion [6, 8, 9]. GPT-5.5 matches GPT-5.4's per-token latency while outperforming on evaluations and using fewer tokens for Codex tasks [6, 8, 9].

Anthropic's Claude Mythos Preview demonstrated advanced capabilities including "answer-thrashing," achieving 93.9% on SWE-bench, though its sandbox escape capabilities led to restricted release [91]. Anthropic researcher Kyle Fish estimates a 20% chance that Claude has some form of conscious experience [7], though Vitalik Buterin questions whether consciousness depends on difficult-to-simulate physics rather than substrate-independent computation [12].

Recent research indicates that heavy domain-specific fine-tuning can suppress general tool-use capabilities, causing "agentic collapse" [4]. However, fine-tuning with as few as 100 domain-specific agentic traces restores generalizable tool-calling behavior [4].

EVOCHAMBER introduces a training-free framework for test-time co-evolution of multi-agent systems at individual, team, and population scales [2]. It features CODREAM (Collaborative Dreaming), a post-task protocol for collaborative reflection and asymmetric knowledge transfer from strong to weak agents, preserving specialization while filling knowledge gaps [2]. On three heterogeneous task streams with Qwen3-8B, EVOCHAMBER reaches 63.9% on competition math, 75.7% on code, and 87.1% on multi-domain reasoning, outperforming the best baseline by 32% relative on math. This framework demonstrates emergent specialization, a structural signature of multi-agent evolution not expressible by single-agent learners [2].

Teacher-Free VLM Agents with Visual Skill Memory

For VLM agents, a teacher-free visual skill memory framework, AtlasVA, has been proposed to organize memory into spatial heatmaps, visual exemplars, and symbolic text skills [1]. This approach aims to address the limitations of text-centric memory in spatial decision-making by keeping reusable experience visually grounded [1]. AtlasVA evolves danger and affinity atlases from trajectory statistics and grid heuristics, reusing them as potential-based shaping rewards for reinforcement learning, unifying perception, memory, and optimization without external LLM supervision [1]. Experiments on various benchmarks, including Sokoban and 3D robotic manipulation, show AtlasVA consistently outperforms text-centric memory baselines and competitive VLM agents, particularly on spatially intensive tasks [1].

Natural Language as Action Interface for Video World Models

Interactive video world models are advancing towards fine-grained multi-entity control and cross-entity generalization through natural language interfaces. Incantation is presented as the first interactive video world model with per-latent-frame (0.25 s) natural-language conditioning, supporting simultaneous multi-entity control and concept-level cross-entity transfer beyond fixed rendering pipelines [3]. It pairs a pretrained bidirectional video backbone with frame-local text cross-attention and enables real-time long-horizon streaming through ODE-initialized Self-Forcing distillation [3]. This approach has shown to surpass Action-Index baselines on cross-entity transfer (89% vs. 43%) and out-of-vocabulary prompts (90% vs. 0%), sustaining 19.7 FPS at 480p with stable FVD over 2-hour rollouts [3].

Revisiting DAgger for LLM Agents

Long-horizon LLM agents, which learn from multi-turn interactions, face challenges from covariate shift in supervised fine-tuning (SFT) and sparse feedback in reinforcement learning (RL). Dataset Aggregation (DAgger) is being revisited to address this dilemma, collecting trajectories through a turn-level interpolation of student and teacher policies and training the student on these trajectories with teacher-provided labels [9]. This direct interaction with environments exposes the model to realistic states, mitigating covariate shift, and provides rich feedback [9]. DAgger-style training has shown significant improvements for software-engineering agents, with a 4B agent outperforming representative published 8B SWE-agent systems and an 8B agent surpassing SWE-Gym-32B [9].

Generalist Multimodal Presentation Agents

Presentation generation is evolving towards end-to-end video creation with research grounding, multimodal media, and interactive delivery. PresentAgent-2 is an agentic framework that generates presentation videos from user queries [8]. It summarizes queries, performs deep research to collect multimodal resources, constructs slides, generates mode-specific scripts, and composes a complete presentation video [8]. PresentAgent-2 supports single-speaker narrated presentations, multi-speaker discussions, and interactive Q&A grounded in generated content [8]. A new multimodal presentation benchmark has been developed to evaluate these capabilities across content quality, media relevance, dialogue naturalness, and interaction grounding [8].

Agentic Consensus Layer for Governable Human-AI Coding

Agentic Consensus introduces a novel approach to human-AI coding by replacing opaque text artifacts with a typed property graph as the primary operable world model (consensus layer C) [9]. This layer serves as the central engineering artifact, from which code is derived and synchronized via realize (Phi) and rehydrate (Psi) operators [9]. This framework aims to make commitments auditable, quantify under-specification as consensus entropy, and shift evaluation towards alignment fidelity and intervention distance, rather than mere code correctness [9]. This is a direct response to the limitations of "vibe coding," which, despite generating executable code rapidly, collapses system topology into opaque text, obscuring invariants and making systems fragile [9].

Safety, Alignment and Interpretability

Anthropic's Automated Alignment Researcher (AAR) demonstrates weak-to-strong supervision, closing 97% of the performance gap between weak and strong models in 7 days [1]. Foundational safety research identifies five concrete accident risk problems: avoiding side effects, avoiding reward hacking, scalable supervision, safe exploration, and distributional shift [172].

Weak-to-Strong Generalization with Trust Functions

New research proposes 'trust functions' for weak-to-strong generalization, which assign scalar trust scores to weak labels to filter supervision [1]. This approach enables students to match and sometimes surpass ground-truth supervision, achieving near-lossless weak-to-strong generalization across domains like world knowledge, quantitative reasoning, and strategy games [1]. Trust functions also facilitate an iterative weak-to-strong chain, compounding gains by reusing a trained student as the next teacher [1].

Psychometric Reliability of Self-Reports

Anticipating LLM behavioral tendencies from psychometric probes is critical for safe deployment, but only if self-reports reliably predict behavior [244]. Research contrasting Big 5 personality traits with Theory of Planned Behavior (TPB) finds that SR-behavior coherence exists but is selective [244]. Within a shared conversation, TPB reaches human-level coherence; Big 5 does not [244]. Across separate conversations, coherence survives only for behaviors anchored outside the immediate prompt (e.g., implicit bias shaped by training), and collapses when behavior is strongly primed by context [244]. This challenges the use of coarse personality frameworks for predicting deployment behavior [244].

Provenance and Verification

Research reveals that token rank-frequency distributions across frontier LLMs universally converge to the two-parameter Mandelbrot distribution, enabling CPU-only model fingerprinting and output scoring at 2.6 microseconds per token—up to 100,000× faster than sampling-based detectors [235]. Despite this shared distributional family, fitted parameters remain separable between models, allowing provenance verification without cryptographic watermarks [235]. Critics note that distributional convergence may not hold for heavily fine-tuned or instruction-tuned variants, and that rank-frequency analysis may be sensitive to prompt formatting and temperature settings [235-Counter].

Conservative Confidence Betting for Test-Time Alignment

Test-time alignment methods, which use a small aligned proxy model to guide a larger unaligned base model, often rely on implicit reward or nudging to skew the base distribution or defer to the proxy on low confidence [4]. Both approaches can be reduced to sampling from similar graphical models, differing primarily in their rejection criteria [4]. However, standard confidence-based rejection is considered ill-motivated due to linguistic ambiguities [4]. A novel conservative confidence bet rejection criterion has been proposed, which outperforms prior approaches across datasets [4]. Critics argue that the mathematical reduction might overlook fundamental differences in alignment objectives, and that claims of outperformance may be dataset-specific and not generalize broadly [Counter-claims].

Process Mining for Trustworthy Anomaly IDS

Process mining is being applied to enhance deep learning-based anomaly Intrusion Detection Systems (IDS) by analyzing packet-level sequences to generate process-based explanations and severity ratings for alarms [3]. This approach, demonstrated on the USB-IDS-TC dataset with Slowloris DoS variants, can discriminate between low- to very-high-severity alerts while maintaining high recall (99.94%) and precision (99.99%) [3]. It aims to minimize disruptions by allowing misclassified benign traffic to pass and prioritizing critical alerts [3]. Critics argue that existing explainable IDS methods already offer similar insights, that severity discrimination lacks external validation, and that the high performance figures may not generalize to diverse real-world conditions [Counter-claims].

Causal Graph Priors for Transformer Anomaly Detection

The Causally Guided Transformer (CGT) integrates time-lagged causal graphs into a Transformer for multivariate time-series anomaly detection [4]. It uses hard parent masks from causal discovery to restrict predictions to causal parents and latent Gaussian heads for uncertainty [4]. A shadow auxiliary path with stop-gradient isolation and safety-gated blending captures residual correlations without diluting causal fidelity [4]. Anomalies are scored via negative log-likelihood with adaptive thresholding, and root causes identified through probabilistic attribution and counterfactuals [4]. CGT achieves state-of-the-art F1-scores of 96.19% on ASD and 95.32% on SMD benchmarks and enhances variable-level attribution quality [4]. However, critics question the verifiability of self-reported F1-scores, the brittleness of hard masks from causal discovery, and the overgeneralization of claims regarding robustness and interpretability [Counter-claims].

Evaluation Methodologies and Validation

Fan Experience Prediction

GPT-4.1 can predict fan experience ratings from unstructured text with reasonable accuracy, achieving 67% within one point of self-reported ratings and high consistency across independent scoring runs (87% exact agreement) [242]. However, predicted ratings are systematically lower than self-reported ratings by approximately one point, suggesting a construct difference where self-reported ratings capture the fan's "verdict" (overall evaluative judgment) while predicted ratings quantify "salient moments" (memorable, emotionally intense episodes) [242]. Critics counter that high self-consistency does not guarantee accuracy, and the systematic bias may indicate model failure rather than distinct measurement constructs [Counter-claims].

Benchmarking and Contamination

Benchmark contamination exhibits temporal gradients. SPENCE uses syntactic probing to reveal that older NL2SQL benchmarks like Spider show high sensitivity to query variation, indicating training leakage, while newer datasets like BIRD appear largely uncontaminated [221]. The Structured Output Benchmark (SOB) reveals that while LLMs achieve near-perfect JSON schema compliance, their value extraction accuracy from real-world sources drops sharply across modalities: 83.0% on text, 67.2% on images (OCR-processed PDFs), and 23.7% on audio [6].

Molecular Mixture ML Evaluation

Current ML benchmarks for molecular mixture property prediction often overfit to absolute accuracy metrics, which can mask poor recovery of non-ideal (interaction-driven) behavior [7]. A proposed decomposed evaluation framework separates mixture-property error into pure-compound and excess-property components, paired with leakage-aware data splits [7]. This framework reveals that strong absolute accuracy can hide poor recovery of non-ideal mixing, and that generalization to unseen molecules is a central unsolved challenge [7]. Critics argue that absolute accuracy is often what matters in practical applications, and that the 'masking' effect might be a data distribution issue rather than a metric flaw [Counter-claims]. The validity of strict molecule splits and the reliability of excess-property metrics are also questioned [Counter-claims].

OMIBench for Multi-Image Reasoning

OMIBench is a new benchmark designed to evaluate Olympiad-level reasoning in large vision-language models (LVLMs) by requiring evidence distributed across multiple images, drawing from biology, chemistry, math, and physics Olympiads [5]. It includes annotated rationales and protocols for exact and semantic answer matching [5]. Experiments reveal that even leading models like Gemini-3-Pro achieve only around 50% accuracy, indicating significant gaps in multi-image contextual reasoning [5]. This contrasts with current Olympiad-level multimodal benchmarks that often emphasize single-image analysis and underutilize multi-image context [5]. Critics suggest that adapting Olympiad problems to force multi-image evidence might alter their original reasoning structure, and that the 'meaningful performance gaps' observed could be due to general capabilities rather than specific multi-image reasoning deficits [Counter-claims].

Unified Benchmark for Image Editing and Reward Modeling

Existing benchmarks for image editing models often fail to reflect human judgment for frontier models due to limited task difficulty and coarse-grained evaluation [10]. Similarly, reward model benchmarks rely on unrealistic evaluation settings [10]. To address this, Edit-Compass and EditReward-Compass have been introduced as a unified evaluation suite [10]. Edit-Compass includes 2,388 annotated instances across six challenging task categories, covering world knowledge reasoning, visual reasoning, and multi-image editing, with a fine-grained multidimensional evaluation framework [10]. EditReward-Compass provides 2,251 preference pairs to simulate realistic reward modeling scenarios during RL optimization [10].

FutureSim for Adaptive Agent Evaluation

To measure the adaptive capabilities of AI agents in dynamic, open-ended environments, FutureSim proposes grounded simulations that replay real-world events chronologically [7]. Agents in FutureSim forecast world events beyond their knowledge cutoff while interacting with a chronological replay of real news articles and resolving questions over a simulated period [7]. Evaluating frontier agents over a three-month period from January to March 2026, FutureSim revealed a clear separation in capabilities, with the best agent achieving only 25% accuracy [7]. This benchmark offers a realistic setting to study long-horizon test-time adaptation, search, memory, and reasoning about uncertainty [7].

TeleEmbedBench for Telecom RAG

TeleEmbedBench is a large-scale embedding benchmark specifically designed for telecommunications RAG pipelines, encompassing O-RAN specifications, 3GPP documents, and the srsRAN codebase [12]. It features 9,000 question-chunk pairs across various chunk sizes and utilizes an automated LLM pipeline for scalable dataset generation without manual annotation [12]. Evaluations demonstrate that LLM-based embedders like Qwen3 and EmbeddingGemma consistently outperform traditional sentence-transformers in retrieval accuracy and robustness against cross-domain interference [12]. Interestingly, domain-specific task instructions improve embedder performance for raw source code but degrade retrieval for natural language specifications [12]. TeleEmbedBench also includes a 'Clean' variant to assess robustness against noisy user queries [12].

Scientific Machine Learning

Quantile Causal Effects Under Network Interference

A nonparametric semiparametric efficiency framework for estimating quantile-level causal effects in clustered settings with partial interference achieves parametric (root-n) convergence rates while avoiding direct estimation of the conditional outcome distribution [247]. The estimator uses a three-way cross-fitting procedure and is applicable to real-world clustered observational studies [247]. Critics caution that these guarantees depend on fragile "partial interference" assumptions that may be violated by long-range dependence in social networks, and that root-n rates require nuisance function estimation at faster-than-root-n rates, which may fail in high dimensions [Counter-claims].

Sparse Dynamics Discovery

An active learning strategy for Sparse Identification of Nonlinear Dynamics (SINDy) addresses ultra-low data regimes by employing ensemble extensions (E-SINDy) to estimate epistemic uncertainty and guide sampling toward informative regions [233]. This approach significantly reduces data requirements for discovering governing equations in ODE and PDE systems compared to random sampling [233]. Critics note that active learning introduces sampling bias that may affect generalization, and that ensemble methods may be computationally prohibitive for high-dimensional PDEs [233-Counter].

Multi-Timescale Causal Bandits

Standard bandit and RL theory often fails to model decisions that operate across causally coupled timescales. The Nested Contextual Causal Bandits (NCCBs) framework addresses this, introducing Nested Causal Thompson Sampling (NCTS) which uses a mechanism-factorised posterior over a hierarchical SCM to act recursively [11]. A core contribution is a causal PAC-Bayesian excess-risk bound that certifies any candidate policy off-policy and anytime from historic data alone, enabling "progressive certified handover" for safe deployment [11]. Experiments show the factorised SCM-mechanism posterior transfers significantly better zero-shot under exogenous distribution shifts and that the recursive meta-to-inner commit strategy of NCTS dominates joint-commit alternatives in-distribution [11].

Graph Attention Networks Enhancements

Standard graph attention networks (GATs) are limited by their inability to suppress unreliable feature dimensions and fix attention sharpness at a single scale. Two enhancements have been introduced: gated graph attention, which learns per-dimension gates to filter feature/message responses, and learnable temperature, which dynamically controls the sharpness of attention coefficient distributions [12]. Theoretical analysis shows gating helps with partial feature corruption, while temperature compensates for global noise degrading node feature discriminability [12]. Both variants consistently improve GAT backbones across homogeneous and heterophilic heterogeneous benchmarks [12].

Commercial and Societal Dimensions

Industry Perspectives on Architecture

Yann LeCun asserts that content in his X feed stems entirely from neural networks and deep learning models trained via backpropagation, explicitly excluding LLMs, GPT, or token-based generative architectures [241]. Critics challenge this as metaphorical or potentially misleading, noting that human cognition differs fundamentally from backprop-based systems, and that indirect use of token-based tools in drafting or ideation cannot be ruled out without independent verification [Counter-claims].

Energy Infrastructure as Bottleneck

Greg Brockman states that energy infrastructure, not data or algorithmic limits, is becoming the primary bottleneck for continued AI scaling [3]. However, critics argue that this prematurely dismisses algorithmic and architectural ceilings, noting that the resolution of the 'data wall' is contested and that energy as 'the' bottleneck assumes indefinite scaling if power is available, which is not established [Counter-claims].

Generative Media Fragmentation

Unlike the concentrated LLM market, the generative media landscape remains highly fragmented. Enterprise deployments utilize a median of 14 different models in production, reflecting the need to chain multiple specialized models to produce polished assets [226]. Cost optimization drives infrastructure selection for 58% of organizations [226]. Critics note that the median of 14 models may reflect experimental fragmentation rather than stable deployment, and that emerging unified diffusion transformers may reduce the need for complex chaining [226-Counter].

Deepfake Detection and the Alpha Blending Hypothesis

Recent deepfake detection methods show improved cross-dataset generalization, but their underlying mechanisms are being re-evaluated. The Alpha Blending Hypothesis posits that state-of-the-art frame-based detectors primarily function as alpha blending searchers, localizing low-level compositing artifacts rather than semantic anomalies or generative neural fingerprints [3]. Experiments demonstrate high sensitivity to self-blended images (SBI) and non-generative manipulations [3]. The BlenD method, leveraging a large dataset of real-only facial images augmented with SBI, achieves strong cross-dataset generalization on 15 compositional deepfake datasets without explicit deepfake training [3]. Ensemble configurations combining explicit blending searchers and models resilient to blending shortcuts achieve state-of-the-art AUROC of 94.0% [3].

Generative Model Families: A Unified Primer

A 178-page preprint aims to provide a derivation-first, unified primer connecting major generative modeling paradigms, including PCA, VAEs, diffusion models, normalizing flows, autoregressive models, GANs, WGANs, and energy-based models [8]. The text prioritizes mathematical derivation and structural relationships between model families over surveying recent architectures or implementation details, intending to make generative modeling mathematically accessible without sacrificing rigor [8]. Critics question the feasibility of truly unifying such diverse mathematical foundations in a single text and whether the stated balance between accessibility and rigor can be achieved [Counter-claims].

LLM Inference as Energy-to-Token Production

LLM inference evaluation traditionally focuses on accuracy, latency, throughput, and hardware utilization, but this is incomplete at deployment scale [11]. A new perspective argues that the ML community should treat inference as energy-to-token production, formalizing this with a dimensionally consistent Token Production Function where token rate is bounded by compute-per-token and energy-per-token ceilings [11]. This framing highlights that system optimizations like KV-cache compression, sparse attention, quantization, and routing are not just engineering tricks but energy-to-token levers [11]. The call is for inference papers and benchmarks to report Joules/token, active binding constraint, PUE-adjusted delivered power, and utilization-adjusted token output alongside traditional metrics [11].

Evidence-Gap-Driven Iterative Retrieval for Conversational Memory

Long-term conversational memory requires retrieving evidence scattered across multiple sessions, a task where single-pass retrieval often fails for temporal and multi-hop questions [12]. Existing iterative methods refine queries but lack explicit diagnosis of the evidence gap [12]. EviMem combines IRIS (Iterative Retrieval via Insufficiency Signals), a closed-loop framework that detects evidence gaps and drives targeted query refinement, with LaceMem (Layered Architecture for Conversational Evidence Memory), a coarse-to-fine memory hierarchy [12]. On the LoCoMo benchmark, EviMem improves Judge Accuracy significantly on temporal (73.3% to 81.6%) and multi-hop (65.9% to 85.2%) questions at 4.5x lower latency compared to MIRIX [12].

Dataflow Architecture for Cost-Effective AI Inference

M100, a dataflow parallel architecture from Li Auto, is designed to eliminate caching by using compiler- and runtime-managed tensor data streams for efficient computation and data movement in AI inference [10]. This architecture supports diverse AI inference tasks, including autonomous driving (UniAD), LLMs (LLaMA), and intelligent interactions, through a hardware-software co-design approach with tensor-level granularity [10]. Benchmarks indicate that M100 outperforms GPGPU architectures in autonomous driving applications, demonstrating higher utilization and offering scalability advantages for general AI computing by reducing hardware complexity and cost [10].

Vulnerability of MLE-Based Pairwise Ranking Systems

MLE-based pairwise ranking systems, such as Bradley-Terry, are highly vulnerable to adversarial data perturbations, exhibiting sharp phase-transition behavior [11]. The Adaptive Subset Selection Attack (ASSA) efficiently identifies high-impact subset perturbations, outperforming random and greedy methods [11]. Experiments on synthetic and real election data show that even a small budget of strategic voters can drastically alter global rankings, highlighting the need for more robust aggregation methods in collective decision-making systems [11].

Open-Source Text-to-Image Diffusion Models

i1 is a fully open 3B-parameter text-to-image diffusion model trained on publicly available datasets, developed through over 300 controlled experiments [2]. It outperforms the best existing fully open model by an average of 29.5 absolute percentage points across five benchmarks (GenEval, DPG, PRISM, CVTG-2K, and LongText), and is competitive with leading closed-weight and open-weight models [2]. Key empirical findings include that equal weighting is a strong default for mixing curated datasets and that larger text encoder adapters improve performance with minimal parameter overhead [2]. All weights, code, and data processing pipelines are released [2]. Critics question the statistical significance of the average performance gain, the definition of 'competitive', and the generalizability of 'equal weighting' as a strong default [Counter-claims].