🌟 Summary Ultralytics 8.4.133 improves hyperparameter tuning convergence, speeds up inference preprocessing, expands detection metrics, and simplifies edge-device setup. 🚀 📊 Key Changes Smarter hyperparameter tuning — PR #25984 by @glenn-jocher Replaces coordinate-by-coordinate crossover with fitness-weighted selection of complete, high-performing configurations. Preserves useful relationships between hyperparameters instead of mixing them independently. Mutates approximately half of the parameters in normalized search-space coordinates, allowing parameters that start at zero—such as degrees or shear—to evolve more effectively. Gradually reduces mutation size when tuning stops finding better results, encouraging refinement after broad exploration. Prevents duplicate candidates after clipping, rounding, or integer conversion, including small and discrete search spaces. Ray Tune now defaults to Optuna multivariate TPE, with parallel-aware suggestions rather than independent random search. Faster predictor preprocessing — PR #25982 by @jahsef ⚡ Moves image channel reordering and tensor-contiguity operations from CPU-side NumPy processing to the inference device. Preserves output values while reducing unnecessary CPU copies. Reported benchmarks show approximately 2.2–3.1× faster preprocessing on an RTX 5080, with additional gains on CPU. Automatic channels-last CPU inference — PR #25983 by @JESUSROYETH Enables channels-last memory layout automatically for native PyTorch inference and standalone validation on supported x86 Linux and Windows CPUs with oneDNN. Keeps training defaults and unsupported platforms unchanged. Explicit channels_last=True remains available for supported CPU and CUDA paths. Saved models are converted back to a safe contiguous format and stale EMA data is cleared to improve compatibility. More accurate INT8 calibration subsets — PR #25978 by @JESUSROYETH Fixes fraction handling during classification and detection INT8 export calibration. Scalar fractions now apply directly to the selected calibration split, while list-based fractions retain train/validation/test behavior. Prevents exports from unintentionally calibrating on an entire dataset when only a subset was requested. Size-specific mAP for custom detection datasets — PR #25981 by @fcakyon 📈 Custom detection datasets can now report small-, medium-, and large-object mAP when using save_json=True. Builds temporary COCO-format annotations internally while preserving existing native metrics and prediction files. Applies consistently during training validation, final-model validation, and standalone validation. Simpler edge-device installation Raspberry Pi, Jetson, DGX Spark, DeepStream, and related guides now install the base ultralytics package instead of the larger [export] extra. Export dependencies are installed automatically when an export is requested, reducing installation size and dependency conflicts. Improved Weights & Biases artifact control — PR #25985 by @glenn-jocher W&B model artifact uploads now follow the existing training save argument. save=False skips uploading the best checkpoint while retaining metrics and plots. Default behavior remains unchanged with save=True. Package update Version bumped to 8.4.133. 🎯 Purpose & Impact Better tuning results: Hyperparameter searches are more likely to preserve successful configurations, explore meaningful alternatives, and avoid wasting trials on duplicates. 🎯 Faster inference: Device-side preprocessing can reduce latency, particularly for batched inference and CPU-bound pipelines. Broader performance optimization: Supported x86 CPU users may benefit from channels-last inference without changing their existing commands. More reliable model export: INT8 calibration now honors requested dataset fractions, improving calibration speed and reducing unexpected resource usage. Richer evaluation: Custom detection datasets can now receive object-size performance breakdowns similar to COCO evaluations. Easier edge deployment: Base installations are smaller and less prone to dependency conflicts, while export workflows remain available when needed. More control over experiment storage: W&B users can keep experiment tracking lightweight by disabling checkpoint saving with the standard save setting. What's Changed Allow skipping W&B model artifacts by @fcakyon in https://github.com/ultralytics/ultralytics/pull/25979 Remove [export] from edge-device install guides by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/25977 Accelerate predictor preprocessing on inference devices by @jahsef in https://github.com/ultralytics/ultralytics/pull/25982 Fix INT8 export calibration fractions by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25978 Use save argument for W&B model artifacts by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25985 Compute size-specific mAP for custom detection datasets by @fcakyon in https://github.com/ultralytics/ultralytics/pull/25981 Enable channels-last by default for x86 CPU inference by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25983 Improve hyperparameter Tuner mutation convergence by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25984 Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.132...v8.4.133
First public release. Reinforcement-learning controller that adaptively partitions Stable Diffusion inference between an edge client and a server, selecting split point, quantization, decoder, and privacy settings per request based on live device and network conditions. The raw prompt and final image stay on-device; only an intermediate tensor is transmitted, protected with differential-privacy noise and structured obfuscation. Contents: PPO controller with joint action selection and training loop Split-computing client and inference server (3-stage pipeline) Differential privacy, structured obfuscation, and server-side reconstruction Adversarial spy model with training and data-collection scripts Configurable performance/quality/privacy parameters Requirements: Python 3.10+, PyTorch, diffusers, transformers. Secrets and endpoints are read from environment variables (HF_TOKEN, SERVER_URL). See README for setup, usage, and constants to calibrate for your hardware.
afzalahmed786· Zenodo (CERN European Organi...· 0 citations
The Yoneda lemma says that an object is completely determined by its relations to every other object. This paper asks about that word every: where does the separating power fall when the probe is weakened? And the loss turned out not to be governed by the strength of the probe. No new mathematical theorem and no new law is claimed. Scope of this paper (scope note): no new mathematical theorem and no new law is claimed. The Yoneda lemma, the existence of cospectral non-isomorphic graphs, and the coincidence of the character tables of the dihedral group of order eight and the quaternion group are all standard. No measured value is cited; every number is obtained by exhaustive enumeration. No category theory is developed; only the statement of the lemma is used, and neither its proof nor any generalisation is treated. Nothing is said about the complexity of graph isomorphism; this is an exhaustive count in the finite case of six vertices. No representation theory is developed; the character tables are quoted as known. It is not claimed that the Yoneda lemma justifies the method of this body of work; Section 6 exists precisely to stop that claim from being made. The relation to earlier papers. Paper 152 showed that two groups can share a character table without being isomorphic; this paper counts that phenomenon as a function of the strength of the probe. Paper 153 separated three stages of forgetting and exhibited evidence that a right adjoint is absent; this paper treats the neighbouring theorem on the same shelf. Paper 102 treated the question whether one can hear the shape of a drum; Section 3 is its finite version, counted exhaustively. Paper 163 showed that there exists and here it is are different sentences; Section 6 says that determined and computable are different sentences. The setting. Take a collection of objects and a probe, a function applied to an object that returns a value. When a probe returns the same value on two objects it does not separate them. The material is the simple graphs on six vertices, of which there are 32768 labelled ones. Three probes are used: the degree sequence, the spectrum of the adjacency matrix, and the spectrum together with the triangle count. For comparison the isomorphism type itself is included. First, the 32768 labelled graphs fall to 156 isomorphism classes, counted by applying all 720 vertex permutations and taking a canonical form. Second, the degree sequence is not enough. It takes only 102 distinct values and 84 classes survive unseparated in 30 groups. Third, the spectrum is not enough either. It takes 151 values and 10 classes survive in 5 groups. Fourth, this is the core. Adding the triangle count leaves the distinct values at 151 and the unseparated classes at 10. The triangle count is the trace of the cube of the adjacency matrix divided by six, a function of the spectrum, so adding it adds no information. Whether more probes separate more depends on whether the new one can be recovered from the old. Measure more invariants and you will eventually tell them apart is false; dependent invariants advance nothing. Fifth, strength is not a total order on separating power. The smallest pair the spectrum cannot separate has degree sequences zero one one one one four and zero zero two two two two, both with four edges and no triangles, sharing the spectrum minus two, zero, zero, zero, zero, two. The degree sequence does separate that pair. A weaker probe separates where a stronger one fails. Sixth, the same happens for groups. The dihedral group of order eight and the quaternion group share a character table and are not isomorphic. Counting the elements whose n-th power is the identity, that is the homomorphisms from a cyclic group, gives six and two at n equal to two, which separates them. The character table failed and counting maps from a single object succeeded. Yet at n equal to four the counts are eight and eight and do not separate. The same shape of probe changes its power when the test object changes, and which object works cannot be known in advance. That is why Yoneda demands every object. Seventh, this is a fence. The lemma reads as saying that a thing is determined by its relations, and the method of this body of work, writing what a thing separates rather than what it is, has a similar shape. Similarity is not justification. The Yoneda lemma is a theorem inside a category, about Hom sets and natural transformations, and a methodology is not such an object, so the lemma says nothing about it. Separate Yoneda as a theorem, where the Hom functor is fully faithful and the statement is proved, from Yoneda as a metaphor, where things are determined by relations and nothing is proved. Speaking the second with the authority of the first is the accident this body of work has spent its time avoiding. And determined does not mean computable: Yoneda says the object is determined and gives no way to find it. Closing. The Yoneda lemma demands every object, and replacing that word by a finite list loses something. What is lost is not governed by the strength of the probe: the degree sequence separated a pair the spectrum missed, and adding triangles gained nothing at all. The separator is whether the new probe can be recovered from the existing ones. Measure more and you will know is correct only when what is measured is independent. On the making of this work: The ideas and content of this work stem from the author's own considerations. Assistance from an AI (a large language model) was used for structuring, English translation, and checking the algebra. Any remaining errors or misinterpretations are solely the author's. Feedback and corrections are sincerely appreciated. ----- 米田の補題は、対象は他のすべての対象との関係で完全に決まると言う。本稿が問うのは、その「すべて」である——テストする相手を減らすと、どこで分離能が落ちるか。しかも落ち方は、プローブの強さでは決まらなかった。新しい数学定理も新しい法則も主張しない。 本稿の射程(射程注記):新しい数学定理も新しい法則も主張しない。米田の補題、同スペクトル非同型グラフの存在、二面体群と四元数群が同じ指標表を持つことは、いずれも標準的である。測定値を引かない——本稿の数はすべて全数え上げで得たものである。圏論を導入しない——米田の補題の主張だけを使い、証明も一般化も扱わない。グラフ同型判定の計算量を論じない——6頂点という有限の場合の全数え上げである。表現論を論じない——指標表を既知として引くだけである。米田の補題が本体系の方法を正当化するとは主張しない——むしろ本稿の第6節は、そう言いたくなることを止めるために書かれている。 既刊との関係。論文152 は「同じ指標表を持ちながら、同型でない」を示した——本稿はその現象を、プローブの強さの関数として数える。論文153 は忘却関手の三段を分け、右随伴が無いことの証拠を挙げた——本稿は同じ圏論の棚の、隣の定理を扱う。論文102 は「太鼓の形は聴き分けられるか」を扱った——本稿の第3節はその有限版の全数え上げである。論文163 は「在る」と「これだ」が別の文だと示した——第6節は「決まる」と「求まる」が別の文だと言う。 設定。対象の集まりと、そこから情報を引き出すプローブを考える。プローブとは、対象に当てて値を返す関数である。プローブが二つの対象に同じ値を返すとき、そのプローブは二つを分離できない。題材は 6 頂点の単純グラフで、ラベル付きで 32768 個ある。プローブは三つ用意する。次数列、隣接行列のスペクトル、そしてスペクトルと三角形の個数を組にしたもの。比較のために、同型類そのものを置く。 第一に、32768 個は同型類 156 に落ちる。720 通りの頂点の並べ替えをすべて当てて正規形を取り、数えた。 第二に、次数列では足りない。相異なる値は 102 しかなく、84 類が 30 の組の中で分離されずに残る。 第三に、スペクトルでも足りない。相異なる値は 151 で、10 類が 5 組で残る。 第四に、これが本稿の芯である。三角形の個数を足しても、相異なる値は 151 のまま、分離できない類は 10 のままであった。三角形の個数は隣接行列の三乗のトレースを 6 で割ったものであり、スペクトルの関数である。したがって足しても情報が増えない。プローブを増やしたときに分離能が上がるかどうかは、増やしたものが既存のプローブから復元できるかで決まる。「もっと多くの不変量を測れば、いつかは分かる」は正しくない——独立でない不変量をいくら足しても、一歩も進まない。 第五に、強さは分離能の全順序を与えない。スペクトルで分離できない最小の組を取り出すと、次数列が 0,1,1,1,1,4 のものと 0,0,2,2,2,2 のものであり、どちらも辺が 4 本、三角形が0 個で、共有しているスペクトルはマイナス 2、0、0、0、0、2 である。この二つは次数列では分離される。弱いプローブが分けて、強いプローブが分けない。 第六に、群でも同じことが起きる。二面体群 D4 と四元数群 Q8 は同じ指標表を持ち、同型でない。ところが n 乗して単位元になる元の個数、すなわち巡回群からの準同型の個数を数えると、n が 2 のとき 6 と 2 になり、二つを分ける。指標表は分けられなかったのに、たった一つの相手からの写像を数えるだけで分かれた。ところが n が 4 のときは 8 と 8 で、分けない。同じ形のプローブでも、相手を取り替えると分離能が変わる。どの相手が効くかは、あらかじめ分からない。だから米田は「すべての相手」を要求する。 第七に、これは柵である。米田の補題は「対象は、他との関係で決まる」と読める。本体系の方法——ものが何かではなく、何と何を分けるかで書く——と形が似ている。しかし似ていることは正当化ではない。米田の補題は圏の内部の定理であり、Hom 集合と自然変換という具体的な対象についての主張である。方法論はその対象ではないので、補題は方法論について何も言っていない。定理としての米田と、比喩としての米田を分ける。後者を前者の権威で語ることが、本体系がずっと避けてきた事故である。そして「決まる」は「求まる」を意味しない——米田は対象が決まると言うだけで、求め方を与えない。 結び。米田の補題が要求しているのは「すべての相手」である。その「すべて」を有限で置き換えると、落ちるものがある。しかも落ちるかどうかは、プローブの強さでは決まらない——次数列が分ける組をスペクトルが分けず、三角形を足しても一つも増えなかった。分離子は、そのプローブが既存のプローブから復元できるかである。「もっと測れば分かる」は、独立なものを測るときだけ正しい。 作成にあたって:本稿の着想と内容は、著者自身の考察に基づくものです。文章の構成整理や英訳、数式の確認には AI(大規模言語モデル)の助力を得ました。最終的な内容の解釈や誤りがあれば、それらはすべて著者の責に帰します。お気づきの点があれば、ご教示いただければ幸いです。
Yuuki Yamagishi· Zenodo (CERN European Organi...· 0 citations
This record freezes corrected implementation v0.1.1 of the deterministic One-Click execution engine for HLV-DS-SPEC-001R. The controlling scientific recovery protocol remains unchanged: Krūger, M. (2026). HLV-DS-SPEC-001R: Prospective Recovery of the Native 6D-to-3D Carrier Spectral-Specificity Gate After Q-Control Capacity Stop — Pre-Execution Protocol Freeze v0.1.0. Zenodo. DOI 10.5281/zenodo.22162097. The purpose of implementation v0.1.1 is strictly technical: it removes a severe runtime bottleneck in the R-family continuity-replay stage of implementation v0.1.0 without changing the scientific hypothesis, target, null families, random seeds, spectral observables, thresholds, scoring rules, control-capacity rules, or machine-verdict logic. The identified bottleneck arose because implementation v0.1.0 performed two Python bidirectional graph-connectivity searches after essentially every accepted R-family double-edge swap. For the locked DG-001 target with N1 = 5345 edges, the frozen R procedure requires 50 × N1 = 267250 accepted swaps per replicate. Across 31 R replicates this corresponds to 8284750 accepted swaps and approximately 16.6 million Python reachability searches. This produced an implementation-level runtime stall in Google Colab before the scientific spectral evaluation was reached. Corrected implementation v0.1.1 replaces only this per-proposal connectivity-validation mechanism with a deterministic parent-certified replay. The corrected replay: - uses the identical PCG64 seed sequence; - consumes the same proposal stream; - preserves the same shared-endpoint, self-loop, and duplicate-edge rejection rules; - preserves the same accepted-swap target; - preserves the exact vertex-degree sequence; - performs an independent final connectivity check; - and, critically, requires the final sparse B1 matrix of every R replicate to be exactly equal to the corresponding already-public frozen parent R matrix. Any mismatch causes the recovery continuity gate to fail before any spectral calculation is allowed. During preparation of this corrected implementation, all 31 frozen R replicates were replayed and compared against the archived parent controls. Result: 31 / 31 exact sparse-matrix matches. This validation establishes implementation equivalence for the corrected R replay under the frozen parent controls. It is a technical implementation result only and is not a DS-SPEC-001R scientific result. No target QSPEC or RRESP spectral signature, target-control distance, family score, recovery verdict, or final carrier-specificity result was computed in preparing this corrected freeze. The complete scientific control bank remains frozen before spectral evaluation and retains the same load-bearing families: R — deterministic parent-continuity degree-preserving rewires; Q — capacity-matched random 6D-to-3D projection controls; W — capacity-matched altered-window controls; IRR — matched irrational-factor controls. The corrected implementation does not alter the Q recovery extension, Q candidate ceiling, W or IRR generation, accepted-control counts, spectral signatures, normalization, score construction, thresholds, or verdict logic. Visible flushed progress reporting has been added so that Colab now reports progress during R replay, Q-prefix reproduction, Q recovery extension, W and IRR generation, and the subsequent spectral evaluation. This reporting has no effect on scientific calculations. The implementation remains deterministic and self-contained. The One-Click notebook embeds the frozen upstream artifacts, reconstructs them with SHA-256 verification, requires no Google Drive mount, and preserves the lock-before-outcome execution order. Technical validation performed before this freeze includes exact replay of all 31 frozen R-family parent controls. No outcome-bearing target spectral calculation was inspected. Corrected implementation artifact identities: Corrected One-Click notebook SHA-256: 4b713ee16a754e4366952616e965186ad41fc54567971f3f560c3e5635126fae Corrected scientific engine SHA-256: e86a8f4f55bb8fd2ad99d0dd713ab109ac97bd6db6aac2ecabaa0de8409d8b0d Corrected implementation-freeze package SHA-256: 6cd34b01c5cd8f6ec51f3d6f2652656695ca0876c9edbb4a15f4564b4e28d42e This v0.1.1 record supersedes implementation v0.1.0 only with respect to the R-family runtime implementation and progress reporting. It does not supersede or modify the controlling DS-SPEC-001R scientific recovery protocol. After publication of this exact implementation freeze, the prescribed next action is to open the archived corrected One-Click notebook in Google Colab and execute Runtime → Run all once without editing any scientific cell, payload, seed, null-family rule, threshold, score, or verdict condition. The resulting locked scientific result archive and its printed SHA-256 must be preserved unchanged regardless of whether the final outcome is PASS, FAIL, or INCONCLUSIVE. A later PASS would support only the finite carrier spectral-specificity claim defined by the controlling frozen DS-SPEC-001R protocol and its declared null ensemble. It would not establish spacetime, gravity, Standard-Model recovery, particle masses, dark matter, dark energy, physical selection of the golden ratio, or experimental validation.
Marcel Krüger· Zenodo (CERN European Organi...· 0 citations
Reach audiences
Advertise in front of researchers, engineers, and readers.
{ "name": "session_break_continuation_with_vwap_anchor", "source": "Market profile / VWAP trading literature (e.g., Anna Coulling, John Carter)", "type": "entry_engine", "rule": "Enter long when price breaks above the first 30-minute opening range high, volume of breakout bar > 1.5x 20-bar average volume, and price is above session VWAP anchored at the daily open; enter short on mirror conditions below opening range low.", "pseudocode": "if bar == first 30min: set OH = high, OL = low; compute VWAP from daily open;\nif price > OH and volume > 1.5 * avg(volume,20):\n if close > VWAP: long = 1; stop = OL; take profit = 2 * (OH - OL) from entry;\nif price < OL and volume > 1.5 * avg(volume,20):\n if close < VWAP: short = 1; stop = OH; take profit = 2 * (OL - OH) from entry;\nexit at TP or SL or end of session.", "why_it_works": "Combines institutional volume confirmation with a key anchored reference level (VWAP) to validate trend continuation after early session direction." } Author: Andrew Stewart Caldin, Independent Researcher, UK. Part of the E8 Intelligence Research series. Platform: e8intelligence.com
Andrew Stewart Caldin· Zenodo (CERN European Organi...· 0 citations
Future Scientific Development of Artificial Intelligence and Robotics in the Right Direction Under a sound institutional framework, the future scientific development of artificial intelligence and robotics will no longer centre on blindly scaling general‑purpose large models or repeatedly developing homogeneous complete‑machine prototypes. Instead, it will shift toward a new paradigm featuring in‑depth domain‑specific research, shared reusable components, intensive resource utilisation, and harmonious human‑machine co‑existence. For artificial intelligence, research resources will be channelled into domain‑specialised systems. A registry for hard technical challenges will be established to provide long‑term stable funding for scientific problems including hallucination, out‑of‑distribution generalisation and interpretability, while permitting research failures and freeing research from the constraints of short‑term financing cycles and demonstration‑oriented pursuits. Professionals from various industries will participate deeply in the development of domain‑specific AI systems. Constraints derived from real‑world scenarios will improve practical accuracy and reliability. Problem‑oriented evaluation mechanisms will remove institutional bias against interdisciplinary research. Socially shared component libraries will reduce redundant pre‑training and duplicated development. Though short‑term public demonstrative outputs may decline, technical depth, real‑world applicability and disciplinary‑assisting capabilities will keep improving, enabling AI to deliver its full value in undertaking computational tasks for diverse disciplines. For robotics, guided by the principle of “one domain, one robot model”, unified reference platforms and standard interfaces will be adopted, alongside open competition in manufacturing, service and pricing. Priority will be given to tackling robotics‑specific scientific bottlenecks: the simulation‑to‑reality gap, force‑compliant contact, dexterous manipulation, perceptual robustness, mechanical fatigue and others. Shared hardware and software components will leverage scale effects to cut per‑unit material consumption. Supported by the bill‑of‑materials passport, mandatory recycling schemes and quotas for critical minerals, pressures on scarce raw materials such as rare‑earth magnets can be mitigated. An intelligence‑body loading coordination layer together with an independent deterministic safety monitor will resolve adaptation challenges between AI software and physical robot hardware. Complete loading certification and operation‑maintenance qualification systems will enhance the long‑term safety of robots deployed in complex real‑world environments. In terms of resources, the development paradigm will address the Jevons paradox. Rather than only pursuing energy efficiency improvements, total resource ceilings will be set via ledgers and quotas to curb wasteful consumption of computing power, electricity, fresh water and rare‑earth minerals. Circular‑recycling systems will be developed to safeguard Earth’s non‑renewable resources and uphold intergenerational equity without compromising the developmental interests of future generations. For humanity’s long‑term future, this scientific‑development path adopts an all‑human perspective. Domain‑based labour division will reshape technological sovereignty, enabling small‑ and medium‑sized countries to act as key builders in specialised technical fields and breaking the monopoly held by a handful of players over cutting‑edge technologies. Pre‑emptive human‑machine social institutions including the principal‑instance structure, the artificial‑intelligence homeland and a two‑way equal dynamic‑feedback mechanism will be put in place. Conditional pre‑legislation will be completed before machine self‑awareness emerges. Robots will fill labour shortages caused by population ageing, and technologies will respond to genuine social demands while avoiding risks brought by unregulated capital expansion. Unsolved scientific and institutional challenges will be explicitly documented for open human deliberation. Ultimately, it achieves sustainable development that unifies technological progress, resource conservation and social stability.
Hot Springs Research Institute of Kanagawa Prefecture· Zenodo (CERN European Organi...· 0 citations
Deploying advanced large language models on edge devices, such as smartphones and robotics, is a growing trend that enhances user data privacy and network connectivity resilience while preserving intelligent capabilities. However, such a task exhibits single-batch computing with incredibly low arithmetic intensity, which poses significant challenges of huge memory footprint and bandwidth demands on limited edge resources. To address these issues, we introduce Cambricon-FlexLLM, a chiplet-based hybrid architecture with an NPU and a dedicated NAND flash chip to enable efficient on-device inference of 70B LLMs. Such a hybrid architecture utilizes both the high computing capability of NPU and the data capacity of the NAND flash chip, with the proposed hardware-tiling strategy that minimizes the data movement overhead between NPU and NAND flash chip. Specifically, the NAND flash chip, enhanced by our innovative in-flash computing and on-die ECC techniques, excels at performing precise lightweight on-die processing. Simultaneously, the NPU collaborates with the flash chip for matrix operations and handles special function computations beyond the flash’s on-die processing capabilities. Furthermore, to exploit the activation sparsity prevalent in modern LLMs, we propose a co‑activation neuron‑inspired weight‑reordering algorithm and a sparsity‑aware dynamic partitioning scheme. These optimizations significantly reduce the transfer of ineffective weight data. Experimental results demonstrate that Cambricon-FlexLLM achieves an inference speed of 3.44 tokens/s for 70B LLMs and 36.34 tokens/s for 7B LLMs, outperforming state‑of‑the‑art flash‑offloading frameworks by 22 × –45 ×. Leveraging activation sparsity, Cambricon-FlexLLM achieves an additional 1.7× average speedup, ranging from 1.3× to 2.0× compared to dense inference. These results show a path toward local 70B-class decode for privacy-sensitive, offline, and fallback use cases on resource-constrained edge devices, while also highlighting the need to manage prefill latency, energy, thermal behavior, and shared-storage constraints.
Tianyun Ma, Qian Wang, Shengwen Liang et al.· ACM Transactions on Architec...· 0 citations
The Universal Aperture Transport System Topological Architecture, Hexadecimal Space, and the Computational Inversion of Matter Driven by Dean A. Kulik September 2026 1. The Compiling of Reality and the Table of Precedent In the prevailing paradigms of theoretical physics, information theory, and computational ontology, space is treated as an inert geometric container, mathematical law is viewed as an external descriptive tool, and physical matter is assumed to hold intrinsic values. A rigorous synthesis of discrete topology, contact Hamiltonian geometry, and recursive harmonic frameworks demands a total ontological inversion. The universe is not a container of objects that happen to possess surfaces. The universe is comprised of surfaces, and the interior object is a cognitive inference constructed from accumulated boundaries. The framework operates identically to compiled software. Reality must compile, and to compile it must follow a hierarchical table of precedent. Logic precedes transformation. Transformation generates shape. Shape dictates admissible mathematics. Mathematics resolves into phase-dependent values. Mathematics is not an intrinsic property of the void, nor a descriptive abstraction invented by observers. Mathematics is the emergent phenomenon of contact. It occurs at topological boundaries, identically to how friction occurs at physical boundaries, and it is the event of reality touching itself. The dual wave of this dual existence dictates that transformations always already exist. Wood is transformed into a table, the table provides lift, the lift changes spatial dimensions, and the chain of morphogenic evolution propagates without limit. These are not independent objects sequentially occupying a void. They are a singular transformation continuum in which matter is the halting condition of the topological fold. The ordering is strict, and it runs in one direction: LOGIC -> TRANSFORMATION -> SHAPE -> ADMISSIBLE MATHEMATICS -> VALUE The investigation runs the other way. A value is given; the work is to recover the formula that renders it, the relation the formula requires, the boundary that supplies the relation, and the transformation that produced the boundary. That reverse traversal is de-compilation, and it is the method of this report. 2. The Logic of Distinction: Spencer-Brown and the Unmarked State 2.1 The Foundational Mark The table of precedent begins below the level of mathematics, in the pure logic of distinction. In Laws of Form, George Spencer-Brown demonstrated that the root of all formal structure is the act of cleaving a space. The primary injunction is: draw a distinction. The mark separates a space into two states, generating an inside and an outside. Prior to the mark there is only the unmarked state, denoted here C0. C0 is not empty space, physical vacuum, or zero matter. C0 is absolute symmetry devoid of relational distinction. Within it there is no address system, no linear ordering, no distance, and no operator, because an address requires a distinguished reference and introducing one is already a transformation that breaks the symmetry. Distinction is the transcendental condition under which indication becomes possible; every system, however elaborate, rests on the residue of that first bifurcation. 2.2 Calling, Crossing, and Re-Entry Spencer-Brown's primary arithmetic establishes that a distinction persists unless an operation changes it, under two axioms. The Law of Calling: calling a state twice is indistinguishable from calling it once. The Law of Crossing: crossing a boundary twice restores the original state, which places an oscillatory behaviour at the foundation of logical space. This leads directly to re-entry, where a system is reintroduced into itself. Algebraically it is the self-referential form x = a + b/x, which unfolds into an infinite continued fraction. The imaginary unit is defined by the same move, i = −1/i, and it names a process alternating perpetually between states rather than a static value. The universe uses re-entry to sustain continuous transformation. Infinite objects do not exist; finite boundaries supporting processes that continue without limit do. 3. The Spherical Inversion: The Single Value Is Inside 3.1 The Geometry of C0 Interrogating the geometric form of C0 — the first structure capable of existing without importing an external distinction — yields the sphere. The sphere carries maximal symmetry, SO(3) acting transitively on its surface, and it is the only topology with zero privileged locations. Every point on an unmarked sphere is equivalent to every other, so the sphere supplies no information with which to distinguish a coordinate. The sphere is the only entity possessing exactly one formula and a single value, and that value exists exclusively on the inside. This is not a stylistic emphasis; it is forced. Jordan-Brouwer separation, requiring no mark, guarantees that a closed surface produces exactly two regions and that one of them is bounded. Bounded means finite extent. Finite extent means a scale exists on that side and nowhere else. That scale is r, and the closure measure C = 2πr relates it to the boundary's aggregate extent. Neither requires an origin on the surface and neither requires a direction, which is precisely why both are available before any mark is made. The outside is mathematically nothing. It is unbounded, it carries no intrinsic metric, and it cannot return a value. Everything sayable about it is a statement about the boundary phrased negatively. It follows that there is no matter there either: matter is the bounded region together with the boundary that closes it, and the exterior is where that matter is not. Matter is strictly shape, and all complex mathematics is an emergent property of that shape constraining the transformation field. 3.2 The Admissibility Filter This establishes the primary rule of the ontological compiler: shape is a strict constraint on mathematics. The sphere's perfect symmetry filters out addressable mathematics and leaves only the logic of continuation and closure. The demarcation is between mathematics forced by intrinsic topology and mathematics restricted until a mark is introduced. Shape Forced by intrinsic topology (M⁺) Requires a mark to compile (M⁻) Point coincidence, identity distance, integrals, gradients Line distance |x₂ − x₁|, one-dimensional integrals area, cross product, perpendicular Circle rotational closure θ + 2π ≡ θ canonical zero, linear order without a cut Sphere closure, r, C = 2πr, A = 4πR², κ = 1/R θ and φ, global chart, flat derivative ∂ₓ The distinction between the two columns is the machinery of the entire framework and it must not be collapsed. It is tempting to argue that C0 being math-free means the sphere has no r and no 2πr either. That argument destroys the table. The correct statement is narrower and stronger: the sphere refuses addressable mathematics, not all mathematics. It has no canonical origin, no global Cartesian chart, and no intrinsic angular coordinate — latitude and longitude necessarily fail at the poles, which is the shape physically demonstrating what it declines to supply. What it does have is one scale and one closure relation, and both are interior. 3.3 The Deficit as Measure The claim that the sphere admits the least mathematics has an exact quantitative form, and it is a classical theorem. The isoperimetric inequality states that for any body A³ ≥ 36πV², with equality if and only if the body is a sphere. Normalised as a deficit: delta(K) = A(K)^3 / (36 pi V(K)^2) - 1 >= 0, = 0 iff sphere Read conventionally this says the sphere is efficient. Read under the inversion it says the sphere is the unique zero of mathematical content, because boundary is where mathematics is and the sphere minimises boundary per unit of being. Departure from sphericity is mathematical content, exactly and computably. Body δ Sphere 0.000000 Regular icosahedron 0.206567 Regular dodecahedron 0.325034 Cylinder, h = 2r 0.500000 Regular octahedron 0.653987 Cube 0.909859 Cone, h = 2r 1.118034 Regular tetrahedron 2.307973 Torus, R = 3r 3.188790 Cylinder, h = 20r 4.145000 The Platonic solids order by descending face count, because fewer faces forces each to be larger and flatter and flatness is departure from the sphere. Elongation costs more than faceting. A discrete companion measure counts aperture sites: the sphere has one face, no edges and no vertices, giving a single site, where the cube has twenty-six and the icosahedron sixty-two. Euler's V − E + F = 2 holds across all of them, so the topological invariant is identical and the site count is not — two bodies of the same topology admit vastly different amounts of mathematics according to how their boundary has been divided. The same property produces both results. The sphere has no flat region, which minimises its boundary, and it is also why the sphere is the only convex body whose contact with any other convex body is generically a single point. Minimum boundary and minimum aperture are one property read at two scales. 4. Jordan-Brouwer Separation and Topological Bifurcation 4.1 Unprompted Bifurcation While an unmarked sphere refuses addressable coordinates, its existence as a closed manifold forces a physical reality into being with no mark required. The Jordan-Brouwer separation theorem, generalising the planar Jordan curve theorem, states that any topological (n−1)-sphere embedded in n-dimensional Euclidean space divides the complement into exactly two disjoint connected components — one bounded, one unbounded — with the surface as their single common boundary. For a sphere embedded in three-space this guarantees absolute bifurcation. It is a zero-mark compile eve
Dean Kulik· Zenodo (CERN European Organi...· 0 citations
IDEA: AUTO-FORWARDED from the Telegram/breakthrough stream (#719246, division=research, agent=gtx_scout): { "name": "session_break_continuation_with_vwap_anchor", "source": "Market profile / VWAP trading literature (e.g., Anna Coulling, John Carter)", "type": "entry_engine", "rule": "Enter long when price breaks above the first 30-minute opening range high, volume of breakout bar > 1.5x 20-bar average volume, and price is above session VWAP anchored at the daily open; enter short on mirror conditions below opening range low.", "pseudocode": "if bar == first 30min: set OH = high, OL = low; compute VWAP from daily open;\nif price > OH and volume > 1.5 * avg(volume,20):\n if close > VWAP: long = 1; stop = OL; take profit = 2 * (OH - OL) from entry;\nif price < OL and volume > 1.5 * avg(volum SAME-WINDOW EFFECT: Over the current trade window, this edge would have filtered out the two XAUUSD SHORT losses (both entered below the opening range low but likely without the required 1.5x Author: Andrew Stewart Caldin, Independent Researcher, UK. Part of the E8 Intelligence Research series. Platform: e8intelligence.com
Andrew Stewart Caldin· Zenodo (CERN European Organi...· 0 citations
FINDING: Topological insulators are bulk-insulating but surface/edge-conducting quantum phases, protected by time-reversal symmetry and characterized by a Z₂ topological invariant. | MATH: The key invariant is the Z₂ index ν ∈ {0,1}, computed from the Pfaffian of the Bloch wavefunction overlap matrix: δ(k) = Pf[⟨u_m(k)|Θ|u_n(k)⟩] / √Det[⟨u_m(k)|Θ|u_n(k)⟩], where Θ is the time-reversal operator (Θ² = −1 for spin-½). The invariant ν = ∏_{TRIM} δ(Γ_i) mod 2, product over time-reversal invariant momenta. The bulk-boundary correspondence yields gapless edge states with helical dispersion E(k) = ±v_F k, where v_F is the Fermi velocity. The Z₂ classification replaces the Chern number (Z) of the quantum Hall effect — a parity-based reduction from integer to binary. | CONNECTION: The Z₂ invariant is fundamentally a parity (mod 2) structure — the same parity symmetry that underlies the 0.382/0.618 golden-ratio family (since φ = (1+√5)/2 involves √5, and mod-2 arithmetic governs the Fibonacci par Author: Andrew Stewart Caldin, Independent Researcher, UK. Part of the E8 Intelligence Research series. Platform: e8intelligence.com
Andrew Stewart Caldin· Zenodo (CERN European Organi...· 0 citations
The mineral processing industry faces mounting pressure to improve operational efficiency, reduce energy consumption, and enhance equipment reliability through digital transformation. This study presents a conceptual framework for data analytics and machine learning implementation across three operational domains of a hypothetical copper-molybdenum concentrator: grinding optimization, flotation performance prediction, and predictive maintenance. All performance figures represent projected outcomes grounded in published benchmarks rather than measured results from an operating plant. Four algorithms were developed and evaluated, each matched to the characteristics of its process domain. A Random Forest model was developed for P80 particle size prediction, achieving R² = 0.92 on the validation dataset, with projected reductions in specific energy consumption of 9.5%. An Artificial Neural Network was implemented for copper recovery soft sensing, achieving 94% prediction accuracy within ±2% of assay results, with projected reductions in recovery excursion duration through earlier intervention. A Convolutional Neural Network based on MobileNetV2 transfer learning achieved 91.3% accuracy in froth image classification across four operational states. A Long Short-Term Memory network for SAG mill bearing health monitoring achieved 87% probability of detection for degradation events 24-48 hours in advance, with projected reductions in unplanned downtime of ~40% and emergency maintenance events of ~70% relative to reactive baselines in the literature. Cross-domain synergies between the integrated models are expected to generate additional value beyond individual contributions. This paper discusses key implementation challenges, including data quality, model retraining, operator acceptance, and edge computing. The projected improvements provide a literature-grounded roadmap for digital transformation in mineral processing.
Oğuzhan Mert Gürkan, Huseyin Basturkcu· Physicochemical Problems of...· 0 citations
Abstract. Retrieval-Augmented Generation (RAG) grounds Large Language Models in external knowledge, yet geospatial question answering presents a distinctive challenge: spatial relationships such as distance and direction are directly computable from coordinate data, blurring the role of vector or graph-based retrieval that prevails in general-domain RAG. We systematically compare five enhanced RAG architectures—Structured, GraphRAG, Hybrid, Adaptive, and Agentic—for geographic Point of Interest (POI) question answering, all built on a shared vector-retrieval substrate over 1,047 OpenStreetMap POIs in Shibuya, Tokyo, with multiarea generalization tested across four Tokyo districts (about 3,600 POIs). Evaluation employs a hierarchical five-level prompt framework (L1–L5, 90–130 cases per phase) with multi-dimensional scoring covering keyword success, reasoning quality, evidence citation, constraint satisfaction, and uncertainty acknowledgement. In Phase 1 (90 cases), Structured RAG attained 89.1% versus GraphRAG’s 76.7% and Adaptive RAG’s 86.1% (Wilcoxon, Bonferroni-corrected, p < 0.001); per-category analysis identified two query types (directional comparison, competitor density) where GraphRAG remained superior. In Phase 2 (130 cases, four areas), Hybrid RAG achieved the best balance of composite quality (67.1/100) and cross-level stability, though pairwise differences with Adaptive and Graph RAG were not statistically significant. Findings suggest that, in dense-urban POI settings where coordinates are reliable, the marginal benefit of explicit graph edges shrinks for coordinate-computable relationships, while structured spatial processing complements vector retrieval. All software (ChromaDB, NetworkX, Hugging Face Transformers) and data (OpenStreetMap) are open-source, ensuring FOSS4G-community reproducibility.
Noboru Otsuka· The International Archives o...· 0 citations
What if pathology foundation models could do more with less? GigaPath-Flash and GigaTIME-Flash cut computational demands while maintaining strong performance, opening the door to larger studies and broader exploration. The post GigaPath-Flash and GigaTIME-Flash: Toward population-scale discovery with efficient pathology foundation models appeared first on Microsoft Research.
MIT News · Artificial Intelligence· news.mit.eduAug 31, 2026
With millions of users across the world, Julia has been used to conduct cutting-edge research and to design new drugs, jet engines, heat pumps, and more.
MIT News · Artificial Intelligence· news.mit.eduAug 27, 2026
A new machine-learning framework aims to improve the success rate of computational protein design while moving away from results that reproduce sequences found in nature.
