August 26, 2026

On Aug 25, 2026, Sber announced GigaChat 3.5 Ultra passed a 250-question test from the 'Information Security' professional-retraining program, scoring 14% above the pass threshold. Test was administered by Sber's cybersecurity unit and the Informzashchita training center. Topics covered Russian information-security law, technical protection against leaks and unauthorized access, cryptographic methods, information-system certification, and InfoSec best practices. Sber says the model was trained on 300,000+ InfoSec documents (best practices, legislation, vulnerabilities, attack tactics and techniques) by Sber's experts. Marketed as a 'first-line consultation resource' for businesses without a dedicated InfoSec specialist and an orientation aid for those entering the profession. Adds to a streak of standardized exam passes: MPEI 'Electrical Power Engineering' and 'Thermal Power Engineering' (May 2026, grade 'good'), Samara State Medical University 'Rheumatology' (Jun 2025), Privolzhsky Research Medical University 'Pediatrics' / 'Neurology' (Apr 2025), Volgograd Medical University 'Cardiology' (Feb 2025), Gnessin Academy 'Musicology' entrance (May 2025).

World model for enterable generative media that responds to continuous navigation while producing 720p video with synchronized environmental sound, music, and speech. Interaction is organized around camera intent: first-person scenes use the camera to define observer motion; third-person scenes learn camera-character dynamics from data. Both discrete commands and continuous poses are mapped to a shared metric-scale relative 6-DoF trajectory with dataset-level motion calibration. A complementary data engine and progressive training scheme plus autoregressive post-training support joint learning of audio-visual generation and trajectory control. Authors report strong trajectory following and visual quality on public world-model benchmarks with both first- and third-person interaction.

OraRL introduces a decoupled advantage estimator for RL post-training of video MLLMs. Policy rollouts set an oracle-free baseline; the oracle-policy gap modulates a directional gain and a separate detached oracle advantage. Sign-balanced pruning keeps only the oracle and strongest rollouts per sign, so the method needs only 2.2x the step time of SFT versus 4.9x for GRPO+CoT. Scales from 0.8B to 9B and up to 100k prompts. Resulting Video-ORA-9B decodes in 130 ms without CoT (vs 4,780 ms), raises temporal mIoU from 62.5 to 66.0, tracking AO 73.0 to 78.2, segmentation 64.3 to 70.4, and scores 73.1 on VSI-Bench versus 55.0 for GPT-5 and 55.1 for Gemini-3-Pro.

Omni-modal model for live-commerce that maps image, video, audio, and text into a unified representation space. Per-vGrid organizes timestamped tokens grouping each video grid with its temporally corresponding audio within explicit boundary tokens for temporal alignment. Three-stage supervised training progresses from omni-modal perception to instruction-following responses. A Faithful-RFT reinforcement stage scores final responses directly with task-verifiable feedback rather than reasoning-style rollout exploration. A scenario-oriented atomic-capability taxonomy and compact data production engine convert live-commerce streams into training signals for ASR, speaker analysis, product visual grounding, OCR, temporal grounding, video dense caption, and omni-modal QA. A synchronized length-grouped sampler reduces padding while preserving workloads across workers; a dynamic sampling strategy regenerates rollout groups with near-zero reward variance to maintain meaningful GRPO advantages.

Identifies two gaps in adapting text-to-image training to image editing: insufficient attention to edit-concept granularity and training inefficiency from sparse supervision. Builds a hierarchical taxonomy of over 1,000 fine-grained edit concepts and a 12M-pair editing dataset (ConceptEdit-12M) via an improved synthesis framework that rectifies distribution collapse while preserving fidelity. Dense supervision synthesizes multiple non-interfering concepts into single image pairs for richer learning signals. Releases ConceptEdit-Bench for granular real-world evaluation.

Interactive, stateful, tool-centric benchmark for mobile planning agents, closing the gap between GUI-centric benchmarks (surface-level screen manipulation) and static function-calling benchmarks (offline API matching). Spans 13 functional domains and 212 realistic mobile tools, running on an executable sandbox with live databases and structured feedback. Evaluates three advanced dimensions: sub-agent collaboration, memory usage, and skill usage. Extensive experiments show current frontier LLMs remain unreliable in mobile settings — performance drops sharply under strict tool ordering, permission limits, and unexpected runtime errors.

Frames harness improvement as an offline learning problem and iteratively updates the harness from failure signals in mini-batches. Combines failure-trace diagnosis, structured patch generation treating the harness as code, and validation-based update selection. Across GAIA2 / SWE-Bench Pro / Terminal-Bench 2.0, gains are 9.0 / 9.6 / 10.0 percentage points respectively. Ablations isolate three effective ingredients: deep debugging rather than shallow reflection, targeted rather than unconstrained edits, and generalization-aware rather than trajectory-specific selection.

DiffusionOPSD converts image-level reward guidance into explicit targets for clean-output predictions at sampled queries, separating target construction from finite realization. A frozen behavior policy generates trajectories and supplies query states and anchors; reward gradients construct bounded positive and negative targets around each anchor; the trainable policy fits these as detached supervision via finite fitting before an EMA refresh of the behavior policy. Achieves best final held-out scores in 19 of 20 reward-matched settings across SD 3.5-M and Z-Image-Turbo and ten evaluators, outperforming the strongest competitor by up to 44%, with 40%/63% training-GPU-hour reductions versus DiffusionNFT.

Formalizes prefix invariance and introduces a lightweight two-forward-pass audit with no training or gradients, producing a per-layer score that pinpoints where causality breaks. Across 192 injected-fault trials on eight checkpoints, mask inspection detected zero, while the proposed audit localized all 192 to the exact layer. Static and dynamic analysis of chunked-scan code in transformers surfaced the same defect in Zamba2 and Nemotron-H, traced to an inter-chunk axis error corrected via the reference implementation.

On Aug 25, Anthropic announced a $5M grant program for independent researchers building open-source evaluations of how AI affects user wellbeing. Grantees receive funding, model access, and technical support; Anthropic requires that evaluations clearly define what they measure, involve clinical/subject-matter experts, test both precautions and harms, reflect real usage (including multi-turn conversations), and validate graders against real experts. Applications due Sep 21; full proposals by Oct 5.