Color coding β tag each field/answer:
- π΅ TECH β technical content
- π’ COMMERCIAL β commercial / business / org content
- βͺ ADMIN β administrative / legal / personal
1. Your Personal Information βͺ
| Field | Value |
|---|---|
| Salutation (Ms. / Mr. / not specified) | |
| Title (Dr. / Prof. / Prof. Dr.) | |
| Last name | |
| First name | |
| Phone number | |
| Your position (e.g. project manager, CEO) | |
| Institution (dropdown) | |
| Institution (other) | |
| Legal Form (None / Ltd. / GmbH / UG / GbR / KG / AG / Individual Enterprise / Other) | |
| Legal Form (other) | |
| Address line 1 | |
| Address line 2 | |
| Postal Code | |
| City | |
| Country | |
| Country (other) | |
| Did you know about SPRIND before learning about the Next Frontier AI Initiative? (Yes/No) | |
| Where did you first learn that SPRIND exists? (dropdown) | |
| Please specify where you heard about us. | |
| What made you decide to actually apply to this Challenge? (dropdown) | |
| Please specify why you decided to actually apply to this Challenge. |
2. Your Solution
Project Title (text, 50 chars) π΅
Recurrent Intelligence
Short Description (textarea, 500 chars) π΅
Total goal: order-of-magnitude efficiency gains and real-time inference. Our bet is the state-based recurrent network β constant per-step compute, no growing context window, memory carried in the evolving state. Low SWaP unlocks dynamic mobile/edge processing: pattern detection in critical-system surveillance and dynamic corrections that prevent failures. The core mechanism we develop β stateful, path-dependent routing with content-addressed memory β is validated by benchmarking and falsification.
Frontier Dimension (textarea, 500 chars) π΅
Model architecture (in-scope: state-space / alternative architectures). A state-based recurrent network opens a scaling axis transformers lack β per-step compute need not grow with model size, and context is unbounded (a fixed evolving state, no re-attention). Built by combining proven components (SSM/Mamba, MoE routing, expressive neurons, recursive reasoning) with new mechanisms we research; existing optimised modules (attention/MLP) used where they help.
Core Idea & Architecture (textarea, 3000 chars) π΅
Substrate. Recurrence is the substrate: a fixed-size state evolves per timestep β streaming-native, constant compute/memory per step, memory implicit in the state (Mamba, S4/S5, xLSTM, HiPPO).
Composition. The architecture is built from classes of proven tools, extended with new mechanisms of our own, composed where each earns its place. The research space we explore:
- Recurrent substrate: selective SSM / spiking / xLSTM-style cells.
- Conditional computation & routing: activate few blocks per step (Sparsely-Gated MoE, Switch), building on routing-by-agreement from capsule networks (Sabour, Frosst & Hinton 2017); incl. stateful / path-dependent routing as one candidate (vs stateless MoE-Mamba/BlackMamba).
- Expressive recurrent neurons: cells richer than a bare leaky integrator (multi-timescale / expressive-neuron class).
- Recursive / hierarchical reasoning: depth decoupled from parameters (GRAM, HRM, TRM).
- Structured sparsity / efficiency: GPU-exploitable, block-structured sparsity.
- Existing optimised modules: attention / MLP, used where they help (hybrid precedent: Jamba, Nemotron-H).
Illustrative candidate (one concrete point in the search space, not pre-committed): a selective-SSM/Mamba backbone carries the recurrent state; each step a stateful router reads that state (hidden + adaptation variables, not just the output) and activates a few expressive, multi-timescale blocks from a larger pool; an optional recursive-reasoning core (HRM/TRM) adds depth without adding parameters; the pool is block-structured so active-only compute maps onto dense GPU matmuls. Compute scales with active blocks per step, not model size; memory stays a fixed-size state.
Which components, which new mechanisms, and how they combine is what the research develops and validates.
Technical Novelty (textarea, 2000 chars) π΅
The core novelty (our bet): stateful, path-dependent routing on a state-based substrate β the active sub-network is chosen from the accumulated recurrent state, not the current token, so the model switches active blocks mid-sequence as context changes, and the routed state doubles as a content-addressed memory. The rebinding: the multi-timescale primitive that ELM/S4/Mamba use inside a unit becomes the gating signal between blocks. Toy evidence: 99% (stateful) vs <70% (stateless) on a cue-switch task.
What current recurrent models miss: hybrids (Nemotron-H, Jamba) still keep attention and route stateless / token-wise (MoE-Mamba, BlackMamba, Routing-Mamba, Swimba); pure SSMs carry a documented recall gap (StateX: 64K-needle 26%β42%). Our delta is to route on the state and address memory by content, aimed squarely at that recall gap.
Why this is reachable, not reproduction: architecture is already out-scaling parameters β GRAM/TRM (~7β10M params) beat a 671B model on reasoning; Nemotron-H beats Qwen-2.5-72B. The innovation is the new mechanism found via disciplined search + invention, not the assembly.
Efficiency as a paradigm: event-driven recurrence + GPU-exploitable structured sparsity define a fundamentally different compute model.
Open hypothesis (to prove/falsify): sub-quadratic compute-vs-sequence-length for the routed/sparse recurrent model β empirical only, no formal claim.
Technical Novelty Citation (textarea, 1000 chars) π΅
Recurrent/SSM: Mamba, Mamba-2, xLSTM, StateX. Routing / routing-by-agreement: Capsules (Sabour 2017), Sparsely-Gated MoE, RMoE, RIMs, Routing-Mamba, Swimba. Expressive neurons: ELM. Recursive reasoning: GRAM, HRM, TRM, Titans. Hybrid at scale: Jamba, Nemotron-H. Neuromorphic delineation: Loihi2, SpiNNaker2, SpikingBrain.
Capability Gap Addressed (textarea, 1000 chars) π΅
The architecture targets pattern recognition in the time domain over a rolling stream of data, with no fixed start/end points β enabling dynamic selection of the data ranges that arise when operating dynamic systems or streaming data. Monitoring dynamic systems continuously is a key control requirement, especially for critical infrastructure and life-threatening medical treatment. Today no model combines unbounded streaming context with frontier-scale capability and efficiency: transformers require a static, bounded input and have a limited attention span; current recurrent / state-space models still rely on limiting training and inference procedures. Our research will establish the conditions under which recurrent networks realise their full potential for dynamic streaming applications, creating the basis for a new class of applications.
Existing Artifacts (textarea, 2000 chars) π΅
Rapid rediscovery of major results β ~11 published theorems/results validated, rediscovered or falsified in ~19 days (11β30 May 2026; ~4 were falsifications of field over-claims). Examples: router state is the critical routing axis (RMoE, Qiu 2024); bare LIF fails long-memory, richer cells recover it (ELM, Spieler 2023); address/content separation governs associative recall (Zoology/MQAR Thm 4.4, Arora 2023).
SSM / Mamba routing (priority artifact). On a continuous diagonal-SSM (Mamba-style) substrate, stateful, path-dependent routing that activates just 1 of 8 blocks (12.5% active compute) reaches 96.7β99.9% on the cue-switch task β vs 47% stateless, 28% random (chance 12.5%), and 99.3% dense (all 8 blocks). Routing nearly matches dense at one-eighth the active compute, and routing on the state is the decisive axis. On harder MQAR associative recall, dense SSM 49% vs routed 25β27% (stateful > stateless > random) β the recall gap is the central open question we target.
Spiking prototype (realizability evidence). ~26Γ inference speedup via a bit-exact LUT over (active block Γ spike pattern); spike-pool kernels 1.73β3.43Γ vs compiled dense, parity-tested (max_abs = 0.0).
Honest caveat. Routed-vs-compiled-dense SSM decode at batch 1 is currently 0.87β0.96Γ β a real routing speedup for continuous SSMs is a kernel research target, not yet realised.
Live artifact dashboard β routed-SSM, kernels, spatial-SNN ablations; versioned benchmark JSONs in the repo. [GAP: GitHub repo + preprint links.]
Technology Readiness Level (TRL) Assessment (textarea, 1000 chars) π΅
Experimental research stage. Validated artifacts exist (GPU kernels parity-tested + benchmarked vs compiled dense; spiking SNN trained on SHD; routing toy) β see Existing Artifacts.
[GAP: formal overall/sub-component TRL numbers β team to assign in next pass; not estimated here.]
| Component | TRL | Justification |
|---|---|---|
| Overall system | [GAP] | Experimental research stage; integrated architecture not yet assembled |
| GPU spike-pool / kernels | [GAP] | Parity-tested + benchmarked vs compiled dense |
| Spiking SNN on SHD | [GAP] | Trained model, accuracy 84β87% @ 20 epochs (ref 90% @ 150) |
| Stateful routing | [GAP] | Toy task validated (99% vs 70%); not yet at scale |
Open Research Risks (textarea, 1000 chars) π΅
- Can a routed/sparse recurrent net match a dense model's accuracy? The central question. A routed SNN collapsed on SHD (V1 38.6% / B1 51.8% vs dense 85.1%). Likely cause: the router decides on binary spike output, so its gradient crosses a surrogate-gradient (Heaviside) boundary and trains poorly. Fix hypotheses to test: (i) route on the continuous membrane potential + adaptation state, (ii) richer block cells than bare LIF (ELM/TC-LIF/PMSN), (iii) key/value separation (CsordΓ‘s).
- Does the inference efficiency transfer to training and to continuous (non-spiking) substrates? Training-time wall-clock speedup unproven for continuous SSMs.
- Is the compute genuinely sub-quadratic in sequence length? Empirical only, no formal proof.
- Does small-scale behaviour survive scale-up? To be checked with scaling curves.
Compute Requirements (textarea, 1000 chars) π΅π’
- Stage-1 estimate [PROJECTED]: validation plus demonstrators to ~13B β 100β180k H100-hours, about 32β64 H100 running continuously over the 7 months.
- Cost β β¬0.3β0.6M at competitive providers β only ~10β20% of the β¬3M envelope: compute is not the bottleneck, personnel and engineering are. Budget a ~2β3Γ reserve for failed runs.
- No brute-force run: a 70B-from-scratch run (~β¬1β1.5M alone) is excluded, on cost and because SPRIND rules out brute-force scaling. Credible path to frontier shown via scaling curves to ~13B.
- GPU type / source: H100-class capacity on competitive cloud / EU providers (exact source to confirm β bulk deal via SPRIND or direct).
KPIs, Benchmarks and Potential Impact (textarea, 1000 chars) π΅π’
Measurement axes (the search criterion = efficiency Γ performance Γ unbounded context):
- Accuracy parity β routed/sparse vs dense at matched FLOPs.
- Latency + throughput (tokens/s decode) vs compiled-dense Transformer & SSM baselines.
- Energy / token; max context length / streaming stability (constant memory over long streams).
- Active-block count / sparsity; compute-vs-sequence-length curve (tests the sub-quadratic hypothesis).
- Specific benchmark suite chosen during the research, on standard long-context / associative-recall / streaming tasks.
Highest impact: real-time / streaming and always-on edge AI (audio, sensor, video, perception, control), infinite-context assistants, and frontier-scale language/code/reasoning under European energy/compute constraints.
Work Plan (textarea, 4000 chars) π’
The project covers research and commercialisation activities across all three stages.
Stage 1 (7 months, β¬3M) β validate the training / kernel approach. Proof with a first model; potential use cases and a partner for co-development identified. Stage 2 (8 months, β¬8M) β scale to larger SNN / recurrent models. Integration tests executed inside co-development projects. Stage 3 (9 months, β¬15.5M) β application / use case. Real-time AI on edge (audio / sensor streams); benchmarks vs Transformer SOTA and neuromorphic hardware; demo of the use case with a partner.
Approach (research method, all stages). Three intertwined modes: (1) build on the field β adapt SSMs, conditional computation, routing / capsule networks, expressive neurons, recursive reasoning, sparsity; (2) original research β develop new mechanisms (state-/path-dependent computation, new recurrent cells, memory and routing schemes, efficient sparse compute); (3) rigorous empirical science β implement β benchmark on the triad (efficiency Γ performance Γ unbounded context) vs a compiled baseline β profile the real bottleneck β try to falsify β keep only what survives. Accelerated by custom AI research tooling (research agents / autoresearch, automated benchmarking, rapid small-model prototyping).
Team / resourcing. Stage 1 is executed by a small, focused team of 5 FTE: the core team full-time plus experts supporting specific technical topics, business operations and go-to-market. Experts may contribute as part-time or temporary hires, subcontractors, or business advisors. For Stage 2 we grow the team, extending specific technical, operational and go-to-market skills β enabling a scale-up of R&D with parallel commercialisation efforts.
Collaboration / subcontracting. Research advisory (Uni LΓΌbeck / S. Otte); co-development partner(s) for use-case integration (Stage 2β3); compute procurement; legal / IP. Specific subcontract work packages defined in the Stage-2 roadmap.
Stage 1 milestones (next research steps):
- M1 β end of month 5: technical report / paper preprint published. Main scope: resolve the central risk β can a routed / sparse recurrent net match dense accuracy? Test the three fix hypotheses (route on the continuous membrane + adaptation state; richer block cells than bare LIF β ELM / TC-LIF / PMSN; key/value separation).
- M2 β end of month 6: artefacts produced (model families, experimental codebase, open-source contributions); a potential scaling dimension or new emergent phenomenon identified, with scaling curves toward ~13B (no brute-force run).
- M3 β end of month 7 (deliverable): updated roadmap for Stage 2 comparing the original hypothesis against the new technological and operational insights. Includes the operational roadmap to scale R&D processes, the growth roadmap to build partnerships and acquire talent, and a detailed financial plan (capital allocation, spending, control mechanisms, cash-flow management).
Financial Cost Estimate for Stage 1 (numeric, max 3,000,000 EUR) π’
Rough estimate (main cost drivers):
| Cost driver | EUR |
|---|---|
| Compute (HW infrastructure & operation) | ~0.3β0.6M (TBD) |
| Personnel (5 FTE) | 500K + |
| Overhead (100%) | 500K + |
| Total | ~1.3β1.6M (within 3,000,000 cap) |
Team (textarea, 2000 chars) π’
Why best suited: the two founders already pursue this thesis (sparsity + recurrence for efficient scale) in published research and shipped production systems; together they cover research, GPU/kernel engineering, and real-time systems, backed by business/strategic and research advisors covering IP, commercialisation and state-based-model science. Real entrepreneurial track record and fast execution. Networks: MIT, Numenta, Mercedes-Benz, Merantix AI Campus, Uni LΓΌbeck, Bitkom, Antler.
Founders / core team:
- Tebjan Halm β 20 years of real-time systems; built a compiler+runtime from scratch; optimised NVIDIA SANA to ~250 ms/image (2β3Γ faster than SANA-Sprint at ~6Γ less compute); state-centric, non-transformer architectures.
- Geoffrey Kasenbacher β spiking/sparse nets & neuromorphic ML at Mercedes-Benz (~1000Γ energy, ~70Γ runtime in a shipped S-Class prototype); 21 granted patents; peer-reviewed (WARP-LCA).
Business & Strategic Advisor: Jana Lehner β physics PhD; Director of IP / ex-CBO at a quantum deep-tech scale-up; 19 years at IBM; Bitkom board; IP moat and commercial/partnerships. Research Advisor: Sebastian Otte β professor at Uni LΓΌbeck, Geoffrey's PhD supervisor; state-based / recurrent models (Active Tuning, recurrent spiking control). Additional researcher (possible): Johann Machemer β DNN pruning research (Calprune); 2 peer-reviewed papers, FLAIRS Best Student Paper; Uni LΓΌbeck.
[GAP: Mirko Klukas (MIT/Numenta, in discussion) and Christian-Hauke Poensgen (fundraising/ops/EU-legal, pending post-defense) β confirm whether to include and in what role.]
What is missing, how we close it: the core is deliberately lean; we deepen hands-on coverage of specific methods/models with targeted senior research & engineering hires as funding grows, drawing on the networks above.
| # | Name, role, % FTE (250 chars) | Track-record links (250 chars) |
|---|---|---|
| Team Member 1 | Tebjan Halm β Founder, [% FTE GAP] | tebjan.de, GitHub |
| Team Member 2 | Geoffrey Kasenbacher β Founder, [% FTE GAP] | GitHub |
| Team Member 3 | Jana Lehner β Business & Strategic Advisor, [% FTE GAP] | [GAP] |
3. Attachments (PDF only) βͺ
- [ ] CVs of key personnel β Jana sends hers tomorrow morning; collect Tebjan + Geoffrey (+ advisors).
- [ ] Detailed cost overview (PDF) β is a SPRIND template provided? If not, Jana can send one tomorrow morning.
- [ ] Declaration on RUS Sanctions form (obligatory) β must be signed by the legal representative.
4. Legal & Declarations βͺ
- [ ] Privacy Notice Agreement β I/we agree to the processing of my/our data in accordance with the SPRIND GmbH (Next Frontier AI Challenge) privacy notice β dated April 2026.
- [ ] Reasons for Exclusion β Infringement Declaration (Yes/No) β not sentenced to >3 months custodial / >90 daily rates / fine β₯ β¬2,500 with an existing non-redeemable Central Trade Register entry (e.g. Β§ 21 MiLoG or Β§ 21 AEntG). Agree to require this from subcontractors.
- [ ] Declaration on Reasons for Exclusion (Yes/No) β no mandatory grounds under Β§ 123 GWB, no optional grounds under Β§ 124 GWB. (If No: proof of self-cleaning per Β§ 125 GWB enclosed.)
- [ ] Declaration of Infringements (Yes/No) β (If No: proof of self-cleaning per Β§ 125 GWB enclosed.)
5. Spam Protection βͺ
CAPTCHA: "What letter is the second last letter of the alphabet?" β Y
6. Submit
β Send SPRIND Challenge submission
Changelog
2026-05-31 22:30 CEST β Prioritise SSM/Mamba artifacts
- Reordered [artifacts]: Existing Artifacts now leads with (1) the rapid rediscovery of ~11 published results/theorems in ~19 days and (2) the SSM/Mamba routing artifact β cue-switch on a continuous diagonal-SSM substrate: stateful routing at 1-of-8 blocks (12.5% active) reaches 96.7β99.9% vs 47% stateless / 28% random / 99.3% dense; MQAR dense 49% vs routed 25β27% (the open recall gap). Spiking prototype demoted to one-line realizability evidence.
- Added [artifacts]: honest caveat β routed-vs-compiled-dense SSM decode at batch 1 is 0.87β0.96Γ (continuous-SSM kernel speedup not yet realised); fixed dashboard links to canonical
/routed/,/kernels/,/spatial/.
2026-05-31 22:05 CEST β Integrate Jana's field texts
- Added [solution]: Jana's drafts integrated into Short Description (merged with the technical bet), Capability Gap (her streaming / dynamic-monitoring framing), Work Plan (her stage structure + 5-FTE resourcing + M1βM3 milestones, with the technical scope grafted in), and the Financial Cost Estimate (5 FTE β¬500K+ / 100% overhead β¬500K+).
- Flagged [capability gap]: softened the "no satisfying model" claim β as written it is easily falsified; reframed to "no model combines unbounded streaming context with frontier-scale capability and efficiency."
- Flagged [financial]: Jana's Stage-1 spend (~β¬1.3β1.6M) is well under the β¬3M cap and contradicts the synthesis-page "compute = 10β20% of the full β¬3M" framing β decision pending.
- Added [attachments / legal]: Jana's notes (CVs tomorrow, cost-overview template question, RUS sanctions + all declarations signed by the legal representative).
2026-05-31 21:50 CEST β Submission form goes live as main entry point
- Added [site]: the submission working-doc is rendered to HTML and published as the site's main entry point (
index.html); technical synthesis moved totechnical-synthesis.html. - Changed [build]: pages now rendered markdownβHTML at build time via
markdown-it-py(no client-side CDN renderer).
2026-05-31 β Form drafted
- Added [solution]: technical fields (Frontier, Core Idea, Technical Novelty + citations, Existing Artifacts, TRL, Open Risks, Compute, KPIs) drafted from the Technical Synthesis page and artifact dashboard.
- Added [team]: founders Tebjan Halm + Geoffrey Kasenbacher; Jana Lehner (Business & Strategic Advisor), Sebastian Otte (Research Advisor), Johann Machemer (possible researcher).