How is AwareLiquid different from Liquid AI LFM?

Liquid AI builds high-performance models using an SSM + linear attention + MoE hybrid for maximum inference efficiency. AwareLiquid builds a continuous-time RNN with neuroscience-inspired modules to study continual learning — can a model learn new domains without forgetting old ones? The research axes are orthogonal: LFM optimises for static-weight inference; MT-LNN explores online weight updates.

Does the needle-in-haystack benchmark work?

Yes, within the base model's context window. Re-tested on 2026-06-26 with the instruct chat template (the earlier 0% was a harness artefact from concatenating raw tokens without the template). Within TinyLlama-1.1B's 2048-token window, retrieval is near-perfect: base 0.867 at 1K and 1.000 at 2K, adapter 1.000 at both. At 4096 tokens both base and adapter drop to 0.000 because that exceeds the base's 2048 RoPE training window — a base-model limit, not an adapter failure. The adapter is at parity with the base in-window; a measurable retrieval uplift remains inconclusive at 5 samples/cell.

What is the MT-LNN API?

The live demo server exposes /v1/completions/stream (SSE streaming) and /v1/model (model metadata). The server is research-grade with no rate limits or SLAs. See github.com/everest-an/M1 for local setup instructions.

M1 · Research

Experiments & Results

Q: Is AwareLiquid production-ready?

No. AwareLiquid is an open research prototype. The 48M from-scratch model produces coherent short stories but is not a general assistant. The MT-LNN adapter generates repetitive outputs at higher temperatures, and a 2026-07 ablation found its own perplexity contribution to be approximately zero over plain LoRA. Both are research artifacts demonstrating architectural ideas.

Q: What is catastrophic forgetting?

When a neural network is trained sequentially on two domains (A then B), its performance on A typically collapses. In AwareLiquid experiments: dense baseline forgetting is +10,562 PPL (244.9 → 10,806.9). EWC reduces this to +1,633 PPL (1,878 final), an 85% reduction. Absolute PPL is still 7.6× original — an open problem.

An honest record of what AwareLiquid M1 has built, what the data shows, and what remains open. Every claim is tagged, reproducible, and falsifiable.

Built Code runs, results reproducible Early Signal Preliminary evidence, stats not conclusive Target Direction clear, not yet achieved

1 · Architecture: AwareLiquid M1

The backbone is the MT-LNN (LTC network) — a continuous-time recurrent architecture whose hidden state evolves according to learnable time constants, rather than discrete layer stacking. M1 (M-Series flagship) layers three cognitive subsystems on top of this core:

Competitive Global Workspace (GWT-B) — selective broadcast of attended information;
Predictive World Model — forward prediction of the next latent state;
Hebbian Plasticity — co-activation-based weight modulation;
Predictive Coding — error-driven representation learning;
Rhythm Modulation — oscillatory gating of information flow.

MT-LNN architecture pipeline diagram: embedding → LTC core → liquid core modules (GWT-B, World Model, Hebbian, Predictive Coding) → output head → continual learning — MT-LNN end-to-end pipeline. The LTC core processes tokens with learnable time constants; liquid core modules add neuroscience-inspired awareness and plasticity; the EWC layer enables continual learning.

Built Architecture, all liquid core modules, and native inference server are implemented and live — this site's demo is driven by them. Technical details: Technical Paper (PDF).

2 · Live Demo: AwareLiquid O1 + M1

The demo offers two distinct models. M1 (the large, cognitive model) wraps a frozen TinyLlama-1.1B-Chat base with a trained residual liquid-core adapter — only a fraction of a percent of the weights are ours (GWT broadcast + EWC continual-learning layers), general language comes from the base. It handles Q&A in English and Chinese and carries persistent episodic memory — session recall tagged with typed cards (identity / preference / emotion / plan / relationship / health / detail). O1 (the small, edge model) is the opposite pole — 48M parameters trained end-to-end from scratch, no base model, the native liquid architecture for hardware / always-on inference, a pure research artifact for short English text continuation.

⚠ Retraction (2026-07): the previously advertised “−28.5% adapter PPL” is withdrawn. A controlled ablation showed that run had the MT adapter frozen by PEFT and trained LoRA only: lora_only 7.984 vs mt_lora 7.920 — the MT adapter contributes ≈0 PPL. The number below measured plain LoRA fine-tuning, not the liquid mechanism. Full detail in RESULTS.md. The M1 adapter’s real, un-retracted value is cross-window / cross-session recall (0.56 vs 0.000), not perplexity.

Metric	Value	Notes
From-scratch params	~48M	Full MT-LNN pre-training
Validation PPL (scratch)	~136	Clean M2 run, 5,000 steps
Adapter PPL gain (M1)	−28.5%	Retracted — was LoRA-only; MT adapter adds ≈0 PPL (7.984 → 7.920)
Inference	CPU	2–4 cores, ~2–15 tok/s

PPL computed with pure next-token cross-entropy — auxiliary losses are excluded to avoid artificially deflated numbers.

Built End-to-end pre-training + native serving. Target Larger, higher-quality model (see demo disclaimer).

Adapter Benchmark Results

The adapter PPL result shown here is retracted (see the note above — it measured LoRA, not the MT mechanism). The context-injection uplift is a prompt-template effect (identical with vs. without the adapter), and the needle-in-haystack retrieval test was re-run 2026-06-26 with the chat template (near-perfect within the base's 2048-token window). We keep the charts below for provenance, clearly labelled.

Bar chart (retracted): Base LLM PPL 9.161 vs 6.553 — the gain was later traced to plain LoRA, not the MT-LNN adapter — **⚠ Retracted.** Fig. 2a showed a −28.5% PPL drop attributed to the MT-LNN adapter, but a controlled ablation traced that gain to plain LoRA (MT adapter frozen, adds ≈0 PPL). Kept for provenance; see the retraction note above and RESULTS.md.

Left: a prompt template that injects the fact lifts accuracy from 83.3% to 96.7% (+13.3pp) — this is the template's effect, identical with vs without the adapter. Right: needle-in-haystack — within TinyLlama's 2048 window retrieval is near-perfect (base 0.87-1.0, adapter 1.0); 4096 collapses for both as it exceeds the base RoPE window. — Fig. 2b — Left: a **prompt-template** that injects the absorbed fact raises QA accuracy by +13.3 pp (30 questions, Qwen-1.5B) — this is the template's effect and is **identical with vs. without the MT adapter**, so it is not an adapter capability. Right: needle-in-haystack re-tested 2026-06-26 with the chat template — within the 2048 window both base and adapter retrieve near-perfectly; the earlier "0%" was a broken-harness artefact, now retracted.

Correction (2026-06-26). The earlier "needle 0% for both variants" claim was a harness artefact, not a model limit: the old script concatenated raw tokens without the instruct chat template, which always scores 0.0 on an instruct-tuned base. Re-tested with the chat template and a faithful adapter rebuild (MT adapter + LoRA, 374/374 tensors verified): within TinyLlama-1.1B's 2048-token window, retrieval is near-perfect — base 0.867 (1K) / 1.000 (2K), adapter 1.000 / 1.000. At 4096 tokens both collapse to 0.000 because that exceeds the base's 2048 RoPE training window — a genuine base-model limit the frozen base never had, not an adapter failure. The adapter is at parity with the base in-window; with only 5 samples/cell and the base near saturation, a measurable retrieval uplift stays inconclusive. Harness: bench_needle_m1_faithful.py.

3 · Continual Learning & Anti-Forgetting

This is AwareLiquid's primary research direction — orthogonal to the frozen-weight paradigm of frontier LLMs. The question: can a model sequentially learn new domains without catastrophically forgetting prior ones?

Experimental Design

Train and evaluate on domain A, then continue training on domain B, then re-evaluate A;
Three arms: dense (standard baseline), liquid (MT-LNN backbone), consolidation (EWC mechanism);
Statistical rigor: paired deltas per seed, SNR = |mean| / std, require all seeds to agree in direction, check "non-catastrophic" threshold.

EWC: Elastic Weight Consolidation

After training on domain A, anchor weights are snapshotted and diagonal Fisher information is estimated from pure CE gradients. During domain B training, a quadratic penalty λ/2 · Σ Fᵢ(θᵢ − θ_Aᵢ)² constrains weights important to A from drifting. This mechanism is architecture-agnostic.

Results (GPU, A = WikiText-103 → B = TinyStories)

3 seeds · 60.7M model · 1,200 steps per domain. Forgetting = PPL rise on domain A after B training (lower is better).

−85%Forgetting reduction vs. dense
early signal — both still forget A

4.14Signal-to-noise ratio
mean / std deviation

3 / 3Seeds with
consistent direction

Bar chart comparing forgetting across three conditions: Dense +10,562, Liquid +9,544, EWC +1,633 PPL rise — Fig. 3a — Forgetting (PPL rise on domain A after B training) by condition. EWC consolidation reduces forgetting by 85% while still learning domain B. 3 seeds, 60.7M model.

Condition	A before	A after B training	Forgetting ↓	Learns B?
Dense (no defense)	244.9	10,806.9	+10,562	✓
Liquid backbone	248.5	9,792.0	+9,544	✓
EWC consolidation	244.9	1,878.0	+1,633	✓

Paired statistics (consolidation − dense): −8,928.8 ± 2,155 · SNR 4.14 · 3/3 seeds. EWC reduces cross-domain forgetting by ~85% without sacrificing domain B learning.

Built EWC mechanism + three-arm experiment framework, fully reproducible (GitHub, experiments/continual_crossdomain_liquid.py).
Early Signal Effect is real, large, and consistent across all seeds (−85%, SNR 4.14) — first clear evidence of mechanism effectiveness.
Not Yet The absolute PPL on A after EWC (1,878) is still ~7.6× its pre-B value (245). Under a strict "≤5× degradation = usable" criterion, this still qualifies as catastrophic forgetting. Verdict: WEAK-TREND, not SUPPORTED.

Honest summary: EWC is working in the right direction and is reproducible — it slashes forgetting by 85%. But "continual learning solved" is far off. Absolute performance on the old domain has not reached the usable threshold. We report this as Early Signal, not a validated success.

4 · Hebbian Plasticity Ablation

The Hebbian co-activation term is one of the MT-LNN liquid core modules. To measure its actual contribution, we ran an isolated ablation: all other v2.0 liquid modules were disabled, only the Hebbian term was active, and hebbian_lr was swept across four values with 2 seeds × 150 steps each.

hebbian_lr	Val PPL (mean ± std)	ΔPPL vs off	Gradient share
off (baseline)	336.9 ± 6.6	—	0
1e-4	337.2 ± 7.2	+0.32	8.3e-8
1e-2	337.2 ± 7.2	+0.32	8.3e-6
1e-1	337.2 ± 7.3	+0.35	8.5e-5

Fig. 4a — Left: validation PPL across Hebbian LR sweep with ±1 std error bars. All values fall within seed noise (grey dashed = baseline). Right: Hebbian gradient contribution on log scale — negligible across all settings tested.

128d × 2L × 8H model · WikiText-103 · CPU · 593s runtime. Largest PPL change (+0.35) is within seed noise (std ≈ 7.3). Gradient contribution peaks at 8.5e-5 — negligible.

Early Signal At current scale, the Hebbian term is effectively inert. Next steps: raise hebbian_lr by 2–3 orders of magnitude or validate at larger model scale before drawing conclusions.

Honest summary: Hebbian plasticity has not produced a measurable effect yet. This is an open experimental direction, not a validated capability.

5 · Spatial Computing (Experimental)

Does MT-LNN's recurrent dynamics suit spatial representation and reasoning? This is a neuroscience-inspired experimental line (grid cells, place cells, path integration). Code is complete and runnable; validation status varies by component.

Implemented Modules

Spatial encoding frontend: grid cell encoding, multi-scale grid modules, head direction cells, boundary distance cells, place cell codes with DoG center-surround patterns — mt_lnn/spatial.py;
Location-indexed associative memory: Hebbian write + pattern completion read, zero trainable parameters — mt_lnn/spatial_memory.py;
Causal spatial steering: detects trajectory ruptures and projects state back onto the legal subspace — mt_lnn/causal_steering.py;
Spatial operators + reasoning scaffold: distance / bearing / reachability graphs, location-based routing — spatial_ops.py / spatial_reasoning.py.

Two Validated Toy Demos

Causal spatial steering: trajectory rupture detected and manifold error projected to 0 — Built Causal steering: an artificial teleportation rupture is introduced at step 18. The steering mechanism detects it and collapses the off-manifold error to 0; the unsteered baseline accumulates error indefinitely.

Spatial memory pattern completion: noisy position query still retrieves correct content — Built Spatial memory pattern completion: noisy position queries still retrieve the correct stored content, and accuracy degrades smoothly — not randomly — as noise increases.

Grid Cell Emergence: Not Yet Validated

MT-LNN and a GRU baseline are compared on a velocity → position path-integration task to check whether hexagonal grid cells emerge spontaneously (standard threshold: grid score > 0.3).

Built Trajectory generation, loss, and grid score autocorrelation analysis are all runnable.
Not Yet All runs to date have been CPU-scale smoke tests. Neither GRU nor MT-LNN has crossed the 0.3 grid-score threshold. Meaningful comparison requires serious GPU-scale training — no spatial advantage is claimed at this time.

Honest summary: the spatial modules and two demo benchmarks are solid and reproducible. But "MT-LNN is better at spatial tasks" has zero data support today. This is an open research direction.

6 · System Benchmarks

Hardware efficiency measurements from reproducible CPU runs. Two key benchmarks: sparse resonance speed–fidelity trade-off, and streaming memory footprint comparison.

Sparse Resonance: Speed vs. Fidelity

MT-LNN uses multi-scale resonance frequencies across 5 time scales. Sparse top-k selection activates only the most relevant scales per token, trading a small divergence from dense logits for significant throughput gains.

Left: throughput bars for dense/sparse k=1,2,3 — sparse k=2 peaks at 6798 tok/s (+24% vs dense). Right: log-scale divergence vs dense — k=2 mean divergence 0.003. — Fig. 6a — Left: throughput by resonance mode (5 repeats, CPU). Sparse k=2 achieves the highest throughput (+24% vs dense k=5). Right: mean absolute logit divergence vs the dense oracle on log scale — sparse k=2 maintains <0.34% mean error.

Streaming Memory Footprint

The recurrent architecture enables a state-only streaming mode that discards historical KV cache after each step, maintaining a constant-size carried state regardless of context length — vs. the O(n) growth of standard KV-cache streaming. On the 832×12 decode benchmark the carried state is 0.381 MB, flat from 512 to 1M tokens (a matched KV-cache reaches 3,072 MB); an earlier, smaller operator-compression config measured 4 KB — the constant is a property of the recurrence, its size a property of the model. Boundary: these are inference carried-state bytes only, not evidence of model quality beyond the 512-token training length — out-of-window LM results are null (see RESULTS.md).

Log-log line chart: matched Llama KV-cache grows linearly from 1.5 MB at 512 tokens to 3,072 MB at 1M tokens, while the O-series carried state stays flat at 0.381 MB. — Fig. 6b — Inference carried state vs. context length, 512 → 1M tokens. KV-cache grows O(n) to 3,072 MB; the carried state stays 0.381 MB. ARR state is a measured snapshot; the KV line is the exact analytic size (832×12, GQA=1). Carried-state bytes only — no quality claim beyond the 512-token training length.

Long-Context Capability (Matched Size)

On Selective Copy — a whole-sequence recall task under heavy noise — MT-LNN leads at every sequence length at matched 200K parameters under a fair full-sequence decode: 0.883 vs 0.672 at T=37, widening to ×2.0 at T=229 (0.219 vs 0.109). The edge is specifically all-or-nothing recall, not average per-token accuracy.

Bar chart, matched 200K params: whole-sequence recall at T=37/101/229. MT-LNN 88%/74%/22%, Transformer 67%/57%/11%, LNN 70%/73%/17%. MT-LNN leads at every length; ratio widens to ×2.0 at T=229. — Fig. 6c — Selective Copy, matched 200K params, equal 1500-step budget, fair full-sequence decode (BENCHMARKS.md). MT-LNN leads at every length; the ratio widens to ×2.0 at T=229. Boundary: toy-scale task — most of the gain is the liquid recurrence; the microtubule machinery adds a smaller increment.

Positioning: Long-Context + Edge Cost, Not AGI Indices

Honest framing: our models (48M–1.1B) score near random on AGI-class benchmarks (MMLU, Agent suites) — those are designed for 50+-point models. We do not compete there. We compete on long-context capability at edge cost: the same-size recall advantage above, plus O(1) carried state (0.381 MB vs 3,072 MB at 1M tokens) — putting per-task cost 2–3 orders of magnitude below cloud APIs.

Log-x scatter: cost per task (USD) vs capability score. DeepSeek-V4-Flash reference at $0.028 / AA index 50. MT-LNN at ~$0.001 / 89.5% (T=32) and ~$0.0009 / 21.9% (T=229, ×2.0). Slay zone shaded. — Fig. 6d — Cost-per-task vs capability, AA slay-line methodology (reference point from the 2026-07-31 Artificial Analysis ranking; costs are estimates). We position on long-context + edge cost, explicitly not on general-intelligence indices.

7 · How to Reproduce

All code, experiment scripts, and reproducible configs: github.com/everest-an/M1
Architecture and pre-training details: Technical Paper (PDF)
Live demo: AwareLiquid Demo · Project overview: About

8 · Cite This Work

If you use this work or build on it, please cite as follows:

BibTeX

@misc{awareliquid2026,
  title        = {{AwareLiquid} {M1}: Bio-inspired Continual Learning with
                  Global Workspace, Sleep Consolidation, and {EWC}},
  author       = {{AwareLiquid Project}},
  year         = {2026},
  howpublished = {\url{https://awareliquid.ai}},
  note         = {M-Series cognitive engine. Open research prototype.
                  Code: \url{https://github.com/everest-an/M1}}
}

APA

AwareLiquid Project. (2026). AwareLiquid M1: Bio-inspired Continual Learning with Global Workspace, Sleep Consolidation, and EWC [Open research prototype, M-Series]. Retrieved from https://awareliquid.ai. Code: https://github.com/everest-an/M1

9 · Changelog

v2.2 June 2026

Brand architecture & legal. Introduced M-Series / O-Series model naming (M1 cognitive slow-thinking, O1 edge fast-thinking). New hero: Beyond Next-Token Prediction. Towards Cognitive Artificial Life. Architecture section rewritten around GWT, NREM sleep consolidation, and Liquid ODE. Engineering Reality data table. Multi-Agent Future section. VS section — Same Lineage. Different Frontier. Privacy Policy + Terms of Use pages. Footer Legal column across all pages. llms-full.txt full sync. JSON-LD dateModified updated.

v2.1 June 2026

Adapter benchmarks & UX. Added Kaggle P100 GPU PPL ablation (−28.5% — later retracted 2026-07: was LoRA-only, MT adapter adds ≈0 PPL), 30-question context-injection uplift (+13.3 pp — prompt-template effect, not the adapter), and a needle-in-haystack retrieval test (re-run 2026-06-26 with the chat template — near-perfect within the 2048 window; the earlier 0% was a broken-harness artefact). New charts: ppl_ablation.svg, context_needle.svg. Dark/light mode toggle across all pages. Scroll-to-top, copy reply, clear chat, retry on offline.

v2.0 May 2026

Full website redesign. Mature production-style layout: sticky nav, hero, stats strip, feature cards, research highlight. Research page with sticky sidebar & IntersectionObserver scrollspy. Dark mode, mobile responsive. Added OG/Twitter meta, canonical, JSON-LD, sitemap.xml, robots.txt, 404.html. Nature-figure matplotlib charts (EWC, Hebbian, sparse resonance, memory scaling, architecture pipeline).

v1.2 Apr 2026

EWC early signal. Three-arm GPU experiment (dense / liquid / EWC) with 3 seeds × 60.7M model. EWC reduces the forgetting delta by 85% (SNR 4.14) — but both arms still catastrophically forget domain A, so this is an early signal, not a validated success. Hebbian ablation: term inert at current scale. Spatial computing section added.

v1.0 Feb 2026

Initial release. MT-LNN v2 architecture: LTC core + GWT-B + EWC + Hebbian plasticity + predictive coding. 48M scratch pre-training on TinyStories. Live demo server. ArXiv-style technical paper.

10 · FAQ

What is the difference between M1 and O1?

M1 (M-Series) is the cognitive slow-thinking engine — it layers GWT competitive broadcast, NREM sleep consolidation with SHY synaptic downscaling, and EWC continual learning on the LTC backbone. It targets episodic memory, agent cognition, and catastrophic forgetting research. O1 (O-Series) is the edge fast-thinking engine — pure continuous-time liquid ODE, constant 4 KB memory state regardless of context length, no KV-cache growth, targets CPU/edge deployment at millisecond latency. Both share the MT-LNN backbone; full cognitive module integration into O1 is in progress.

Is this production-ready?

No. AwareLiquid is an open research framework. O1 (48M from scratch) produces coherent short stories but is not a general assistant. M1 (TinyLlama-1.1B + MT adapter) generates repetitive outputs at higher temperatures, and its MT adapter contributes ≈0 PPL over plain LoRA (see the 2026-07 retraction in §2). Neither is production-grade — both are research artifacts demonstrating architectural ideas. There is no SLA.

How is this different from Liquid AI / LFM?

Liquid AI builds high-performance models using an SSM + linear attention + MoE hybrid for maximum inference efficiency — targeting industrial time-series, edge signal processing, and autonomous driving. AwareLiquid layers cognitive architecture (GWT awareness, sleep consolidation, EWC online learning) on top of the liquid backbone to study continual learning — can a model acquire new knowledge without erasing the old? Same academic lineage (MIT LNN), orthogonal research axes. AwareLiquid is not affiliated with Liquid AI, Inc.

What is catastrophic forgetting?

When a neural network trains sequentially on domain A then B, its performance on A collapses. In AwareLiquid M1 experiments: dense baseline degrades from 244.9 → 10,806.9 PPL (×44). EWC reduces this to 1,878 PPL (×7.6, an 85% reduction in the forgetting delta). The threshold for "usable" retention is ≤5× degradation — not yet achieved. This is reported as Early Signal, not a solved problem.

Can I use the model API?

Yes. The live demo server exposes /v1/completions/stream (SSE, O1), /adapter/v1/completions/stream (SSE, M1), and /v1/model (model metadata). See GitHub README for local setup. Research-grade — no rate limits, no SLAs.

What does the needle-in-haystack test actually show?

Scientific honesty — including correcting our own earlier mistake. We first reported "0% for both variants", but that turned out to be a harness artefact: the old script fed raw concatenated tokens to an instruct-tuned base without the chat template, which always scores 0.0. Re-tested on 2026-06-26 with the chat template and a faithful adapter rebuild, the real picture is: within TinyLlama-1.1B's 2048-token window, retrieval is near-perfect (base 0.867/1.000 at 1K/2K, adapter 1.000/1.000). At 4096 tokens both collapse to 0 because that exceeds the base's 2048 RoPE window — a base limit, not an adapter failure. The adapter is at parity with the base in-window; proving a retrieval uplift needs a larger-window base and more samples. Extending long-context retrieval remains a target for future work.