Wolf Digest — Friday, September 11, 2026

NCP-ArchPreview trains an 8.9B-parameter model on 5.73T Dolma-3 tokens with a Next Concept Prediction objective alongside standard next-token prediction. The model builds a product-quantized concept vocabulary directly from its own hidden states, predicts future multi-token concepts via a dedicated Concept Module, and feeds those predictions back to the token level to guide generation, with both objectives trained jointly end to end.

The efficiency claim is the headline: consuming only 51.3% of the total training tokens, it reaches the final pretraining loss of OLMo-3-7B. After full pretraining it beats OLMo-3-7B by 2.45 points on the downstream macro-average, including a 5.99-point gain on GSM8K. Using 85% of the standard computation it approaches the training loss of a strictly parameter-aligned 8.9B baseline. Controlled experiments separate the contribution of the latent architecture from the NCP objective. The authors describe it as the largest demonstration of a latent-space language model to date, which is the claim independent replication will need to settle.

Most robot policies are Markovian, but many manipulation tasks are not — they need memory beyond the current observation. Existing mechanisms use language summaries or growing visual windows, losing fine-grained visual evidence or trading history coverage against execution efficiency. MaP-WAM splits the problem into memory-grounded planning and plan-conditioned execution, treating long-term multimodal episodic context as planning-time evidence instead of re-conditioning the executor on full history.

Memory is stored as completed segment records holding language instructions and sparse visual context, converted into compact plans of a next segment-level language plan plus visual guidance. A World-Action-Progress model executes each plan over unknown duration, jointly predicting action chunks and execution progress, calibrating progress through plan-observation alignment for adaptive segment transitions. Executor context length stays fixed and structured attention enables KV caching in both phases; the authors report state-of-the-art results on RMBench.

On-Policy Self-Distillation lets a model act as its own teacher using privileged information such as ground-truth solutions, but recent findings show it can badly degrade complex reasoning: forcing the student to imitate an artificially confident trace conditioned on the answer suppresses expressions of uncertainty and penalizes the exploratory, self-corrective behavior hard problems require. Negative Self-Distillation inverts the setup — instead of imitating a privileged solution, the model generates a question-specific negative condition (for instance, prompting itself to act as a careless reasoner) and the student distribution is pushed away from that self-generated negative teacher.

Naively applying unlearning objectives here breaks the model, because flawed reasoning tokens are confounded with ordinary linguistic tokens and penalizing both indiscriminately degrades basic language capability. The fix is a dynamic gating mechanism that identifies reasoning-critical tokens so gradient updates target only those. No ground-truth answers or external supervision are needed, which is what makes the method attractive where verifiers are unavailable.

SenseNova-U1.5 is an 8B mixture-of-transformers model that understands, reasons about and generates visual content inside a single encoder-free and VAE-free architecture — dropping both the separate vision encoder and the latent autoencoder that most unified models still carry. Training uses spatially coherent patch reconstruction to strengthen the visual interface, curated generation and editing data, improved task formulation, structural prompt enhancement, and native resolutions up to 4K.

Post-training optimizes specialized experts for visual aesthetics, bilingual text rendering, infographic generation and image editing, then consolidates them through multi-expert on-policy distillation. Reported gains span image fidelity, text rendering, complex composition, multi-reference editing and interleaved generation, while improving instruction following and preserving subject identity, geometry and unmodified regions. Despite limited exposure to structured formats during generation training, the model generalizes to long structured visual instructions, which the authors read as multimodal understanding transferring into visual planning. Training code for supervised fine-tuning, RL and on-policy distillation is promised as open source.

RTK compresses terminal output before an agent reads it, and has been promoted with claims of cutting Claude Code tokens by 60% and saving 90% of tokens. Quesma benchmarked it on Terminal-Bench 2.1 across 1,740 attempts — Claude Code with Fable 5.0 on 85 tasks and OpenCode with DeepSeek V4 Pro on 89 tasks, five runs per task with and without RTK.

Costs did not fall. Fable 5.0 showed a 5% overall cost reduction but only 1% per-task when weighted equally; DeepSeek V4 Pro got 5% more expensive overall and 17% more per task. Pass rates dropped 1-2%. The tool's own <code>rtk gain</code> metric reported 349.2 million tokens saved across DeepSeek attempts, an 89% reduction, while actual spend rose — one task's two commands accounted for 69% of the reported savings. Three causes: terminal output is only 11% of Fable's input tokens and 40% of DeepSeek's; RTK induced extra agent turns that cost more than compression saved; and one incompatibility produced 339 consecutive errors costing 9x the baseline. RTK's own README already disclaims that cutting bash output is not the same as cutting the bill.

Shanghai Enflame Technology, the Tencent-backed AI chip developer, surged on its first trading day Friday after raising about 6.12 billion yuan ($911 million) in its IPO, with shares opening at 410 yuan. The scale of the pop underscores investor appetite for domestic alternatives to Nvidia inside China.

The listing is a financing event rather than a technical one — no new silicon was announced — but the capital raised and the valuation it implies are the practical measure of how much runway Chinese accelerator vendors now have to iterate against export controls.

Search agents are usually judged on final-answer accuracy, which says nothing about how evidence was gathered. SearchAtlas converts raw search trajectories into structured graphs whose edges record how evidence propagates from the query that retrieved it through to the final answer. The automated parsing pipeline hits a mean edge F1 of 86.0% against human-annotated graphs and stays consistent across repeated runs.

Applied to five search agents across three benchmarks, it exposes fragmented answer support, question constraints that never reach the answer, and unverified parametric knowledge entering responses. These process failures correlate with incorrect answers more strongly than an LLM judge given either the raw trajectory or the ordered query list — the practical argument for structural process analysis over trajectory-level judging.

Two probing methods predict whether an agentic task will succeed from the model's own activations. Latent Trajectory Dynamics summarizes changes in residual-stream representations across an interaction trajectory; the Action Representation Probe predicts success from representations formed at action-decision points.

Across Bash, SQL and Python interactive benchmarks and three model families (Qwen14B, Qwen7B, DeepSeek6.7B), both consistently beat surface-level generation and sequence-based calibration baselines. The deployment argument is that this is a zero-overhead reliability monitor: no prompt alterations, no multi-sample rollouts, just a read off representations the forward pass already computed.

RAND examines how AI middle powers — states that buy frontier systems without hosting the labs that build them — can use public procurement as a governance instrument. The report develops a set of procurement guarantees and assesses whether coordinated purchasing standards across such states could strengthen their bargaining position against vendors.

The mechanism is the interesting part: rather than regulating model development directly, which requires jurisdiction over the developer, procurement conditions attach requirements to the act of buying. The open question the report takes up is whether coordination is achievable in practice, since the leverage depends on buyers not defecting individually.

Neurosymbolic systems lean on solvers for correctness guarantees, but a solver cannot tell whether a formal translation preserves reference-equivalence to the intended formalization. The paper names this Verdict-Preserving-Unfaithfulness: an incorrect encoding that executes cleanly and returns the expected verdict. They prove that structural, verdict-only verification heuristics are mathematically bounded to chance-level detection on such traces — a hard limit, not an engineering shortfall.

Generative Verification distills an offline Z3-equivalence oracle into a reference-free continuous equivalence score computed in the language model's native vocabulary space. The relevance extends past autoformalization: any pipeline that treats "the checker accepted it" as ground truth inherits the same blind spot.

VWG-Bench tests whether video generators can execute symbolic rules, respect physical laws and pursue intentional goals, spanning 9 reasoning dimensions and 38 fine-grained tasks. A three-level VLM-as-Judge protocol separately scores video-level fluency, task-level rule adherence and sample-level goal realization, which is what keeps visual quality from masquerading as cognitive correctness.

Leading models score strongly on rendering and consistently fail logic-heavy and rule-constrained tasks. The proposed remedy, Vid-PRE, is a model-agnostic prompt rewriter trained with RL on purely text-based rewards that offloads the reasoning burden to a dedicated VLM and emits concise constraint-aware prompts.

Omni-modal models ingesting audio, video and text together face KV cache growth that text-only quantization methods handle badly. Analyzing TurboQuant, a rotation-based approach, on multimodal caches surfaces two failures: temporal key drift and heterogeneous value geometry across modalities.

OmniKVQuant is training-free and fixes both — it sets the key quantization range over each short window of the input stream, and rotates values separately per modality. On Qwen2.5-Omni and Qwen3-Omni it enables 2-bit KV caches while substantially preserving performance across seven audio-visual benchmarks. A fused Triton decode kernel unpacks the 2-bit cache during attention so no dense FP16 cache is ever materialized, which is what turns the compression into actual memory savings rather than a bookkeeping result.

Low-bit quantization of the attention module introduces large errors at very low bit-widths. Rather than relying on smoothing to handle outliers, HyQuant quantizes most attention states to low-bit formats while retaining a small set of vertical-line tokens and local-window states in high precision, selecting those accuracy-critical regions with lightweight vertical-line-aware attention-pattern signals.

In prefill, a hybrid-precision quantized attention operator preserves vertical-line tokens and a local sliding window at full precision while quantizing the rest of the context; in decode, the same principle applies to KV-cache compression, fusing dequantization with attention computation. The authors report near-lossless accuracy across tasks, models and datasets with a notably simple design. Code is released.

Compound LLM systems typically solve coordination by adding a higher-level LLM that reads worker outputs, writes the final answer, allocates later calls and decides when to stop — concentrating three control decisions in one opaque, order-sensitive model call. UnitBoost asks whether the manager needs to be generative at all, and replaces it with a defined meta-level operator: a task-given unit map turns worker outputs into slot-value proposals, a constrained argmax assembles the output, and unfilled or unsupported slots become an explicit residual for the next round.

The operator is order-free, records unit provenance, and carries a guarantee that without coupling constraints, unit-wise maximization under the same admission score dominates selection of any complete candidate. On three held-out benchmarks it exceeds the best single candidate chosen with gold labels by 0.060-0.195 absolute task-score points and input-matched generative managers by 0.048-0.076; replacing only the management step improves six compound-system configurations by 0.013-0.182.

Few-step distillation speeds up autoregressive video generation but collapses diversity: under the same prompt, different noise samples yield near-identical videos with weakened motion. The analysis identifies a structured uncertainty collapse specific to the autoregressive setting — DMD's mode-seeking bias maps different noise samples to nearly identical first chunks, and the deterministic AR cache then propagates that collapsed state through every subsequent chunk, turning a local loss of stochasticity at the rollout root into global suppression of temporal variation.

Uncertainty DMD targets that root cause directly rather than adding noise downstream. The diagnosis is the transferable part: it explains why diversity collapse in AR video is qualitatively worse than in single-shot image distillation.

Code world models represent environments as executable programs, but that representation alone does not say how to build a complex world. RCWM reconstructs complex 3D worlds in code from a single reference image by pairing a Recursive Scene Program representation with a construction solver that recursively calls itself.

Each solver call follows the same process — establish the whole, recursively reconstruct unresolved parts, then revisit the whole to refine composition. This global-local-global recursion gives fine-scale structures their own perception-and-editing loops while preserving scene-wide geometry and relationships, which is the standard failure point for compositional scene reconstruction.

Serving cost scales with output length, and preference alignment routinely inflates verbosity without improving utility. LOCUS starts from the observation that low-rank subspaces alter sequence length without modifying the alignment loss itself, then selects a task-aware low-rank adaptation subspace that minimizes output-token cost subject to a utility constraint, retaining the native preference objective with a frozen backbone.

Evaluation uses Anthropic HH-RLHF dialogue preferences across two roughly 3B decoder backbones, Pythia-2.8B and Qwen2.5-3B, against protocol-matched full-parameter DPO and DrDPO branches and the released SamPO checkpoint. Treating length as a property of the update parameterization rather than something to penalize in the reward is the conceptually interesting move.

Datacenter GPU power is the binding constraint on serving capacity, and production has shifted to prefill/decode disaggregation. Deploying NVIDIA's Max-Q inference profile on a disaggregated B200 system produced a modest realized gain of +8.6% tokens per joule, model-dependent, with a mean end-to-end latency cost of +5.2% that throughput-only evaluation does not surface. The profile also applies one setting across prefill and decode GPUs that operate in opposite hardware regimes.

The hypothesis is that the optimal power setting belongs to the deployed combination of model, quantization, engine and hardware rather than to the GPU class, that each lane deserves its own profile, and that converting SLO headroom into energy savings safely requires latency-gated calibration at runtime. The measured latency cost hidden by throughput-only reporting is the transferable warning.

Tool-augmented legal agents produce agentic hallucinations where tool-call and reasoning errors cascade into fabricated holdings and miscited authority. Existing legal benchmarks evaluate single-turn QA with outcome-level metrics, and general agentic-hallucination benchmarks lack legal diagnostic capability; neither answers to what extent and how an agent hallucinates along its trajectory.

LexAgentHallu, built through a four-stage expert-in-the-loop pipeline, contains 3,414 instances across 17 legal categories and 6 task types, structured hierarchically to localize where in a trajectory the failure originates rather than only scoring the final answer.

A multi-signal detection pipeline combining fine-tuned DeBERTa-v3 classification, Monte Carlo Dropout uncertainty quantification and temperature-scaled calibration reaches F1 0.915 and AUROC 0.977 on HaluEval general-domain tasks, with per-task F1 of 0.97 on QA, 0.96 on summarization and 0.82 on dialogue; MC Dropout inference lifts accuracy to 93.2%.

A context ablation supports genuine entailment reasoning over surface pattern-matching — summarization F1 drops 24% when knowledge context is removed — and learning-curve analysis shows 25% of training data captures 77% of full-data performance. Applying DPO to a Qwen2.5-0.5B generator using this detector as the signal cuts its hallucination rate from 85.5% to 37.7%, a 55.9% relative reduction, though measuring the improvement with the same detector that supplied the training signal is the obvious caveat.

Enterprise agents cannot be evaluated on customer production data, and no existing substitute supplies ground truth. This benchmark constructs a complete fictional company — product simulators, company-specific internal databases, benchmark questions and computed answer keys — parameterized by industry, company size, business model, application portfolio and a random seed.

One seeded entity graph feeds simulators of Salesforce, Zendesk, Slack, Gong and other products, while a question-conditioned generator creates schemas and records for internal databases drawing shared entities, keys and values from that same graph. Because the whole estate is generated, answer keys are exact rather than annotated.

ElevenLabs and Universal Music Group announced a multi-year strategic agreement between the audio-AI platform and the largest music rights holder. The post is filed under ElevenLabs' Company category and does not publish licensing scope, artist consent mechanisms, compensation terms or financials, so the operative details — which catalogues, which products, and how artists opt in — remain unstated.

Lambert has compiled and is maintaining a curated list of the best writing on open models from recent years, intended as a comprehensive path for someone getting up to speed on the area. Updated September 11, 2026, with an open call for additions.

The "Waymo effect" is defined as what happens when a technology removes the friction of dealing with another person and the removal is experienced as pure gain. Applied to research, the argument is that LLMs let researchers bypass human collaboration in a locally rational way while losing serendipity and intellectual challenge.

The essay draws on Tim Wu on the tyranny of convenience, Borgmann's device paradigm, Bjork's desirable difficulties, Bainbridge's 1983 ironies of automation, and Dashun Wang's March 2026 Nature commentary, plus team-science findings that small teams disrupt while large teams develop. It presents no publication statistics, collaboration metrics or citation analysis of its own — the case is theoretical, and worth reading as a hypothesis rather than a finding.

Google has released the Gemini app for Windows, moving the assistant from browser and mobile onto the desktop as a first-class application. Distribution rather than capability is the story — desktop placement is the surface where assistant usage competes most directly with incumbent OS-level integrations.

Cloudera and Mistral announced a partnership to deliver specialized AI capabilities inside enterprise data environments, positioned around data sovereignty and integration with existing organizational data systems rather than a model release. It lands the same day as Cohere's sovereignty-framed open-weight translation model, both aimed at buyers who will not send data to a hosted frontier API.