Submission Form

Color coding — tag each field/answer:

• 🔵 TECH — technical content
• 🟢 COMMERCIAL — commercial / business / org content
• ⚪ ADMIN — administrative / legal / personal

1. Your Personal Information ⚪

2. Your Solution

Project Title (text, 50 chars) 🔵

Recurrent Intelligence

Short Description (textarea, 500 chars) 🔵

A model that never stops reading: it predicts the next token over an in-principle unbounded context at constant cost per step, on a device, not a datacenter. We build recurrent networks (RNNs) that hold their whole history in a fixed state, so compute and memory per step stay constant. State-space models are the strongest substrate today; we work across the RNN design space and keep what performs. They scale from tiny sensor streams to frontier LLMs, opening systems as a new modality.

Frontier Dimension (textarea, 500 chars) 🔵

Recurrent networks as the successor to attention-based transformers. Today's models add compute and context, hitting energy and memory walls, hardest under European constraints. A recurrent network carries its history in a fixed state, so it scales to effectively unbounded context and runs on the edge, where transformers cannot. They already win on streaming and long-context tasks. We bet the next S-curve is recurrent; the question is which architecture reaches frontier scale first.

Core Idea & Architecture (textarea, 3000 chars) 🔵

Biological intelligence is recurrent, and the world is continuous: living systems make sense of a stream of events through an evolving internal state, not a fixed block of context. We bring that principle to silicon. The payoff is models that keep predicting over endless streams at a per-step cost that never grows, on the edge rather than in a datacenter.

We develop and optimize recurrent architectures into models that run in the real world and compete in demanding environments. State-space models are the strongest substrate today. The work is to take SOTA recurrent networks and engineer them with the right methods and proper software around them, into fast, efficient, and deployable architectures on standard GPUs, not only on neuromorphic chips.

Several methods drive this. One, our novel invention, is state-conditioned structured sparsity: at each timestep, a stateful router selects a small subset of the model's compute blocks as a function of the accumulated recurrent state and current input, rather than running the full dense model. Another is activation-centric GPU kernels that execute only the active work and map it onto dense matmuls and Tensor Cores, where fine-grained sparsity normally gives no speedup. Another is static structured sparsity, for example structured pruning carried into pretraining, a lever we are developing toward large training speedups, up to roughly 10x as the target. The block pool adds multi-timescale cells and an optional recursive-reasoning module that adds depth without parameters, with attention or MLP used where they help.

A worked example: watching a dynamic system to predict failures. A power grid, a patient, or a production line emits many sensor streams at once. The model folds them into its running state step by step, learns the system's normal dynamics, and detects the hidden patterns that precede a fault, raising a pre-warning before a catastrophic failure is obvious so an operator or clinician can act. The same operator can query the live state at any moment. This shape carries across power grids, health monitoring, or manufacturing surveillance.

The model and router are trained together so the predicted activations stay accurate and stable. We have already built and measured the hard parts: bit-exact GPU kernels that run this sparse compute about 3.4 times faster than torch.compile() dense, and a structured-sparse recurrent model that matches dense accuracy. Benchmarks are in Existing Artifacts. We did all of this in a few days at the Merantix AI Campus in Berlin, leveraging AI research assistance, and we are excited to see what else we can find.

The architecture is not fixed in advance. We run controlled ablations over candidate blocks and their combinations against dense and SSM baselines, profile where the real cost is, and keep the constellations that win; the best mix may be task-specific. Where labelled data for a target system is scarce, we build targeted datasets with design partners.

Technical Novelty (textarea, 2000 chars) 🔵

Our novelty is the set of new mechanisms we develop to solve the distinct limitations of RNNs, composed with proven tool-classes. One is state-conditioned routing with content-addressed memory, where the recurrent state selects the active compute and recalls past content, aimed at the memory limitation. The other is activation-centric algorithms that leverage the network's sparsity, executing only the active work so sparse recurrent compute runs fast on standard GPUs. Further mechanisms target training cost, for example structured pruning carried into pretraining. These are examples; we keep researching where existing architectures fall short.

The combination draws on state-space and recurrent dynamics (Mamba, Mamba-2, xLSTM, StateX), expressive units (ELM), recursive or hierarchical reasoning (HRM, TRM, GRAM, Titans), conditional computation and routing (Capsules, Sparsely-Gated MoE, RIMs, MoE-Mamba, BlackMamba, Routing-Mamba, Swimba), hybrid recurrent-attention precedent (Jamba, Nemotron-H), and neuromorphic roots (Loihi 2, SpiNNaker2, SpikingBrain).

What current recurrent models miss: hybrids such as Nemotron-H and Jamba lean on attention, the routed ones (MoE-Mamba, BlackMamba, Swimba) route stateless and token-wise, and pure SSMs show a recall gap. This is achievable: architecture already beats scale, with small recursive models (GRAM, TRM) beating far larger ones at reasoning, and Nemotron-H beating a 72B transformer.

Measurable advantage today: bit-exact GPU kernels run 3.4x faster than torch.compile() dense at identical accuracy, and structured sparsity matches dense accuracy on a recurrent benchmark. We target order-of-magnitude lower energy per token and large training speedups (estimates). The capability gap is efficient long-horizon inference with reliable memory at bounded, predictable compute, which transformers cannot reach at any scale.

Technical Novelty Citation (textarea, 1000 chars) 🔵

Mamba 2023 https://arxiv.org/abs/2312.00752 Mamba-2 2024 https://arxiv.org/abs/2405.21060 xLSTM 2024 https://arxiv.org/abs/2405.04517 StateX 2025 https://arxiv.org/abs/2509.22630 Capsules 2017 https://arxiv.org/abs/1710.09829 Sparsely-Gated MoE 2017 https://arxiv.org/abs/1701.06538 RIMs 2019 https://arxiv.org/abs/1909.10893 MoE-Mamba 2024 https://arxiv.org/abs/2401.04081 BlackMamba 2024 https://arxiv.org/abs/2402.01771 Routing-Mamba 2025 https://arxiv.org/abs/2506.18145 Swimba 2026 https://arxiv.org/abs/2603.06938 ELM 2023 https://arxiv.org/abs/2306.16922 HRM 2025 https://arxiv.org/abs/2506.21734 TRM 2025 https://arxiv.org/abs/2510.04871 GRAM 2026 https://arxiv.org/abs/2605.19376 Titans 2025 https://arxiv.org/abs/2501.00663 Jamba 2024 https://arxiv.org/abs/2403.19887 Nemotron-H 2025 https://arxiv.org/abs/2504.03624 Loihi 2 2021 https://arxiv.org/abs/2111.03746 SpiNNaker2 2021 https://arxiv.org/abs/2103.08392 SpikingBrain 2025 https://arxiv.org/abs/2509.05276

Capability Gap Addressed (textarea, 1000 chars) 🔵

Always-on streaming at constant per-step compute and memory, where attention and its growing cache cannot follow: live patient monitoring, power-grid and factory telemetry, network and sensor streams. New methods leverage structured and dynamic sparsity on the GPU. Reasoning at small parameter counts, where recursive depth substitutes for size, which suits on-device and regulated settings where data must stay local, such as clinics, banks and law firms. Systems as a new modality: continuous understanding and prediction of a dynamic system, a regime with no fixed start, end, or bounded context. The same architecture can also serve frontier language models and infinite-context assistants.

Existing Artifacts (textarea, 2000 chars) 🔵

These artifacts are the validated tools and R&D machinery we use to build the target system, not the product itself, which is a novel frontier-grade recurrent model showcased by systems as a modality. The items below are proven components, evidence and tooling.

• The core mechanism already works on an SSM substrate. On a continuous diagonal-SSM (Mamba-style) substrate, a stateful router that activates 1 of 8 blocks (12.5% active) reaches 96.7–99.9% on a cue-switch task, versus 47% stateless, 28% random and 99.3% dense. On harder MQAR recall, dense reaches 49% and routed 25–27%, which is the recall gap we want to close.
• We already have methods for dynamic and structured sparsity with real GPU gains. We established which sparsity is hardware-exploitable (block-structured, not fine or random). By predicting which blocks of neurons are active, the model runs extremely fast, and the relative speedup grows with model size: a bit-exact look-up-table kernel rises from about 9x at small scale to about 26x at large versus torch.compile dense, and a wavefront kernel from 1.3x to 5.6x with depth; spike-pool kernels run 1.73–3.43x, parity-tested (max_abs = 0). These kernels go where the gains are needed, inside the recurrent model.
• R&D velocity. We validated, rediscovered or falsified about 11 published results in roughly 19 days, about 4 of them falsifications of field over-claims, for example that router state is the critical axis (RMoE), that only block-structured sparsity is GPU-exploitable, and that structured sparsity can match or beat a dense baseline in accuracy on a recurrent (SNN) benchmark.
• Honest caveat. Routed versus compiled-dense SSM decode at batch 1 is 0.87–0.96x, so a continuous-SSM kernel speedup is still a research target.

Live dashboard: routed-SSM, kernels, ablations. Code and versioned benchmark JSONs are in the project repository (available to the jury on request); a technical report or preprint is in preparation for Stage-1 M1.

Technology Readiness Level (TRL) Assessment (textarea, 1000 chars) 🔵

Overall system TRL 2–3. The concept is formulated and backed by component-level precedents, but the integrated architecture is experimental and needs validation under controlled benchmarks. Sub-components:

• Recurrent / state-space substrate: TRL 4–5. Demonstrated at lab and model scale, with partial evidence of scalable training and inference.
• Conditional routing and sparse MoE: TRL 4–5. Established, but integration with recurrent state-conditioned control is lower maturity.
• State-conditioned routing and content-addressed memory: TRL 2–3. A key research mechanism (one of several), supported by toy-task evidence, not yet validated at scale.
• Structured sparsity and accelerator execution: TRL 3–4. Implementable and hardware-relevant, but dependent on kernel design, batching and routing regularity.
• Hybrid attention and MLP components: TRL 5–6. Mature, reused where ablations show benefit.

Open Research Risks (textarea, 1000 chars) 🔵

The central question is whether we can solve the current limitations of recurrent networks, above all their memory. They compress history into a fixed state and lose precise recall. Starting from architectures like Nemotron-H and Mamba, can we add reliable, content-addressable memory without a growing context window? That is the question that matters most.

Tied to it: can a routed or sparse model match a dense one? Today it does not. A routed spiking net collapsed on SHD (38.6 to 51.8% versus 85.1% dense); we test fixes: routing on the continuous state, richer cells (ELM, TC-LIF), and key/value separation.

Further risks: sparsity may not become a real GPU speedup if routing is irregular; the training speedup from pruning is a target, not proven; small-scale wins may not survive scale-up toward 13B; and labelled data for real systems is scarce.

We test each against dense and stateless baselines, with scaling curves and design-partner data, keeping what survives.

Compute Requirements (textarea, 1000 chars) 🔵🟢

Stage 1 funds validation, ablation and larger demonstrators. We budget about 150,000 to 200,000 H100-hours, roughly 50 to 65 H100 GPUs over the seven-month phase, about 0.4 to 0.5M EUR at competitive cloud or European pricing (an estimate), with a re-run reserve (factor 2 to 3) for failed runs and sweeps. That is about 15 percent of the budget; personnel and engineering are the larger investment.

We establish controlled scaling curves across four to five model families up to about 13B parameters, benchmarked against transformer, SSM, Mamba and hybrid baselines, to prove architectural leverage. Full frontier scale needs an order more compute, secured in Stages 2 and 3 via long-term procurement and SPRIND bulk allocation.

Hardware is H100-class via cloud or European providers, plus a one-time local workstation GPU spend of about 90k EUR for fast per-researcher ablation.

KPIs, Benchmarks and Potential Impact (textarea, 1000 chars) 🔵🟢

KPIs follow efficiency, performance and long-horizon operation: accuracy parity/improvement of routed-sparse vs dense at matched FLOPs; decode latency and throughput (tokens/s) vs compiled dense transformer and SSM baselines; energy per token; max stable stream length at bounded memory; realised sparsity and active-block count; and compute vs sequence length (sub-quadratic test).

Benchmarks include long-context retrieval, associative recall, cue-switching, streaming classification and real-time prediction, final suite chosen in Stage 1.

The lead use case and highest impact is systems as a modality, the continuous understanding and prediction of dynamic systems (power-grid, factory, medical), as always-on, in-house performance on small servers, plus infinite-context assistants and frontier-scale language, code and reasoning under European energy and compute constraints. Impact is assessed by matching benchmarked strengths to use cases, then validating with a design partner.

Work Plan (textarea, 4000 chars) 🟢

The project covers research and commercialisation across all three stages.

Stage 1 (7 months, €3M), validate the training and kernel approach. Proof with a first model, and potential use cases plus a partner for co-development identified. Stage 2 (8 months, €8M), scale to larger recurrent models. Integration tests run inside co-development projects. Stage 3 (9 months, €15.5M), application and use case. Real-time AI on the edge (systems and sensor streams), benchmarks against Transformer SOTA and neuromorphic hardware, and a demo of the use case with a partner.

Approach (research method, all stages). Three intertwined modes. First, build on proven tool-classes such as state-space models, routing, expressive neurons, recursive reasoning and sparsity. Second, do original research to develop new mechanisms, such as state and path-dependent computation, new recurrent cells, routing schemes and efficient sparse compute. A specific bet is to attack the memory and recall problem with feedback cycles (iterative, top-down recurrence) and inference-time selective weight adjustments (test-time weight updates and fast weights). Third, run rigorous empirical science: implement, benchmark on the triad of efficiency, performance and unbounded context against a compiled baseline, profile the real bottleneck, try to falsify, and keep only what survives. This is accelerated by custom AI research tooling for automated experimentation, benchmarking and rapid small-model prototyping.

Implementation and commercialisation. The lead use case is systems as a modality, the continuous understanding, prediction and observation of dynamic systems (power-grid, factory, medical). Each validated approach is benchmarked, matched to use-case requirements, and taken to co-development partners to build one real, validated use case rather than a generic LLM replacement.

Hardware path. GPU-efficient models are the Stage-1 priority. Neuromorphic-chip partnerships are a longer-term lever, and chip-design itself is out of Stage-1 scope.

Team and resourcing. Stage 1 runs on a small, focused team of 5 FTE, the core team full-time plus experts who support specific technical topics, business operations and go-to-market as part-time or temporary hires, subcontractors or advisors. For Stage 2 we grow the team and extend technical, operational and go-to-market skills, which scales R&D with parallel commercialisation.

Business operations from day one. We form a European legal entity (UG minimum) early in Stage 1, budget at least half an FTE for business operations (finance, controlling, HR), use external providers for HR, accounting and IP-law, and include an IP-protection budget and a post-grant follow-up and financing plan. (Operations and Economic-Viability prose: Jana.)

Collaboration and subcontracting. Research advisory (Uni Lübeck, S. Otte), co-development partners for use-case integration in Stages 2 and 3, compute procurement, and legal and IP. Specific subcontract work packages are defined in the Stage-2 roadmap.

Stage 1 milestones.

• M1, end of month 5. Technical report or paper preprint published. Main scope: resolve the central risk, whether a routed or sparse recurrent net can match dense accuracy, by testing the fix hypotheses (route on the continuous membrane and adaptation state, richer block cells such as ELM, TC-LIF or PMSN, and key-value separation).
• M2, end of month 6. Artefacts produced (model families, experimental codebase, open-source contributions), and a potential scaling dimension or new emergent phenomenon identified, with scaling curves toward about 13B (no brute-force run).
• M3, end of month 7 (deliverable). Updated Stage-2 roadmap that compares the original hypothesis against new technological and operational insights, with an operational roadmap to scale R&D, a growth roadmap for partnerships and talent, and a detailed financial plan covering capital allocation, spending, control mechanisms and cash flow.

Financial Cost Estimate for Stage 1 (numeric, max 3,000,000 EUR) 🟢

Submitted figure: 1,345,000 EUR (an estimate, within the 3,000,000 cap; non-dilutive PCP). Direct-cost breakdown below; full version in the cost-overview PDF.

Team (textarea, 2000 chars) 🟢

Why best suited. The two founders already pursue this thesis, sparsity and recurrence for efficient scale, in published research and shipped production systems. They have a proven research track record and high momentum: in only a few days, many results were proven and rediscovered casually on a laptop GPU. Together, they cover research, GPU and kernel engineering, and real-time systems. Networks include MIT, Mercedes-Benz, Merantix AI Campus, Uni Lübeck, Bitkom and Antler.

Founders / core team:

• Tebjan Halm. 20 years of thinking in recurrent systems design, GPU programming, real-time 3D, and theoretical computer science; core developer of vvvv, a VPL with a compiler and runtime with state hot-swap built from scratch; optimised NVIDIA SANA to about 250 ms per image (2 to 3x faster than original SANA-Sprint).
• Geoffrey Kasenbacher. From neuromorphic computing: spiking and sparse nets and neuromorphic ML at Mercedes-Benz (about 1000x energy and 70x runtime in a shipped S-Class prototype); 21 granted patents; peer-reviewed (WARP-LCA).

Business and Strategic Advisor: Jana Lehner. Physics PhD; Director of IP and former CBO at a quantum deep-tech scale-up; 19 years at IBM Quantum; Bitkom board; managed two €10M grants; covers IP, commercial and partnerships. Research Advisor: Sebastian Otte. Professor at Uni Lübeck, Geoffrey's PhD supervisor and second examiner of Johann's bachelor thesis; researches state-based recurrent models. Researcher (in talks): Mirko Klukas. Math PhD; sparse-coding and sequence-memory research at MIT and Numenta. In talks and agreed to join as a hire; not 100% confirmed before the deadline. Additional researcher (possible): Johann Machemer. DNN pruning research (Calprune); 2 peer-reviewed papers and a FLAIRS Best Student Paper.

What is missing and how we close it. The core is deliberately lean; we add targeted senior research and engineering hires as funding grows and outsource HR, accounting and IP-law.

3. Attachments (PDF only) ⚪

• [ ] CVs of key personnel. Jana sends hers tomorrow morning; collect Tebjan and Geoffrey, plus advisors.
• [ ] Detailed cost overview (PDF). Is a SPRIND template provided? If not, Jana can send one tomorrow morning.
• [ ] Declaration on RUS Sanctions form (obligatory). Must be signed by the legal representative.

4. Legal & Declarations ⚪

• [ ] Privacy Notice Agreement. I/we agree to the processing of my/our data in accordance with the SPRIND GmbH (Next Frontier AI Challenge) privacy notice, dated April 2026.
• [ ] Reasons for Exclusion, Infringement Declaration (Yes/No). Not sentenced to more than 3 months custodial, or more than 90 daily rates, or a fine of at least €2,500, with an existing non-redeemable Central Trade Register entry (for example § 21 MiLoG or § 21 AEntG). Agree to require this from subcontractors.
• [ ] Declaration on Reasons for Exclusion (Yes/No). No mandatory grounds under § 123 GWB, no optional grounds under § 124 GWB. If No, proof of self-cleaning per § 125 GWB enclosed.
• [ ] Declaration of Infringements (Yes/No). If No, proof of self-cleaning per § 125 GWB must be enclosed.

5. Spam Protection ⚪

CAPTCHA: "What letter is the second last letter of the alphabet?" answer Y

6. Submit

→ Send SPRIND Challenge submission

Changelog

2026-06-01 — Compute/budget reconciled

• Compute Requirements and Financial fields disagreed on H100-hours; reconciled to ~180k H100-h / ~€450k cloud. Total €1.59M → €1.345M.

2026-06-01 — Clickable citations, de-duplicated fields, story spine, framing fix

• Citations [novelty citation]: now 21 raw clickable https://arxiv.org/abs/... URLs keyed to the [n] markers used in the narrative (plain-text form fields only render raw URLs).
• Framing fix [novelty / core idea]: the architecture is recurrent models; the stateful router is one of several sparsity-leverage methods (with activation-centric GPU compute and structured sparsity), not the core architecture.
• De-duplicated [solution]: each field now does one job. Frontier carries the story spine + S-curve + continuous-AI bet; Core Idea is the architecture/data-flow (approved routing wording); Capability Gap restructured around efficiency / controllability / reasoning / new use-cases.
• Team: backgrounds added (Geoffrey from neuromorphic computing; Tebjan from 20 years of recurrent systems design, 3D and theoretical computer science) plus proven track record and high momentum (results proven/rediscovered casually on a laptop GPU in days).

2026-06-01 — Continuous-AI bet, momentum proofs, memory research bet

• Added [frontier]: the bet that the world is continuous and needs continuous AI, served by a combination of methods rather than one.
• Sharpened [artifacts]: momentum now states two concrete proofs, structured sparsity matching or beating dense accuracy on a recurrent (SNN) benchmark, and predicting which blocks of neurons are active to run extremely fast.
• Added [work plan]: a research bet to solve the memory and recall problem via feedback cycles (iterative, top-down recurrence) and inference-time selective weight adjustments (test-time weight updates, fast weights).

2026-06-01 — Efficiency-first reframe, plain prose, compute scaled up

• Reframed [solution]: thesis is now "we want to solve the long-standing problems of recurrent networks" for frontier-grade, order-of-magnitude efficiency. Systems-as-a-modality dialed back to the showcase use case. Stateful routing kept as one mechanism among several. Attention framed as a tool used where it helps.
• Style [all §2]: removed em-dashes and AI-writing patterns; converted claims to goal framing.
• Scaled up [compute / financial]: (superseded 2026-06-01 below) compute and total were briefly raised; see the compute-rebalance entry.
• Trimmed: all fields now within their character limits.

2026-06-01 — Reframe groundwork + Google-Doc sync + artifact reframe

• Synced [solution]: §2 prose to the team Google Doc; artifacts framed as tools/evidence for the novel recurrent model, not the product. Lead use case set to systems understanding / prediction / observation.

2026-05-31 22:30 CEST — Prioritise SSM/Mamba artifacts

• Reordered [artifacts]: rapid-rediscovery and the SSM/Mamba routing result (cue-switch 96.7–99.9% at 1-of-8 blocks); honest decode caveat.

2026-05-31 22:05 CEST — Integrate Jana's field texts

• Added [solution]: Jana's drafts (Short Description, Capability Gap, Work Plan structure, Financial figures).

2026-05-31 21:50 CEST — Submission form goes live as main entry point

• Added [site]: submission working-doc rendered to HTML as the site's main entry point; build-time markdown to HTML via markdown-it-py.

2026-05-31 — Form drafted

• Added [solution]: technical fields drafted from the Technical Synthesis page and artifact dashboard; founders and advisors set.