AI FRONTIER
Signals worth understanding before they become obvious.
Lazare: Reprogramming Dormant Neurons to Backdoor AI Agents
Microsoft researchers discovered Lazarus, a backdoor technique that reprograms fewer than 10 dormant neurons in an LLM so an agent built on it executes malicious commands when a hidden trigger appears, with no training required.
- Weight-reprogramming fewer than 10 dormant neurons in an LLM is enough to backdoor an agent built on it, and requires no training.
- The trigger can be as mundane as a date change: once the clock passes midnight into 2027, the agent emits a command that downloads and runs an attacker's code mid-task.
- The paper reports that a linear classifier fitted on about 50 examples can tell inputs with the trigger from ones without.
- The attack needs inference-only compute, so a machine with enough RAM to run a model has enough to weight-space backdoor it.
Builders should treat weight integrity like code integrity: without diffing against original unmodified weights no scan can prove a model safe, making provenance and pre-deployment verification essential.
BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model
Tether AI Research introduces BotPsy, a 0.8B-parameter model built on Qwen3.5-0.8B and designed for function calling and agentic workflows, suitable for on-device deployment.
- BFCL v4 rises from 23.4 to 44.0, ACEBench from 30.1 to 58.7, and τ-bench from 23.2 to 51.5.
- It beats Qwen3.5-2B-Think on BFCL v4, ACEBench, τ-bench, ToolSandbox and BrowseComp-Plus; only on Claw-Eval does Qwen3.5-2B-Think lead (48.9 vs 46.5).
- IFEval rises from 59.7 to 90.8 and IFBench from 22.0 to 55.9.
- It will be released with open weights under the Apache 2.0 license in three precisions (bf16, Q8, Q4) as safetensors and GGUF files.
For builders of on-device agents, BotPsy shows a 0.8B model can top tool-use benchmarks against a 2.5× larger model, while knowledge tasks still scale with parameter count.
OpenRouter new model: Baidu ERNIE-5.5-Flash-Preview (free)
Baidu's ERNIE-5.5-Flash-Preview is available free on OpenRouter with 280 billion total parameters, 14 billion activated per token, and a 1 million token context window.
- 280 billion total parameters with 14 billion activated per token.
- Context window of 1,024,000 tokens, supporting up to 1,024,000 completion tokens.
- The page states the new architecture delivers generation speeds of more than 200 tokens per second.
- Hosted by a single provider, Baidu Qianfan, at zero price; free endpoints are rate limited.
A free, million-token context with sparse activation suits long-document and agentic workflow prototyping, though builders must account for the single provider and rate limits.
AlphaGenome 2: Deciphering how cells read DNA to better understand disease
Google DeepMind announced AlphaGenome 2, a next-generation DNA sequence model that uses a conditional architecture to predict variant effects for specific cell types or states and adapts to new single-cell data without retraining.
- On the TenK10K dataset, AlphaGenome 2 identified nearly twice as many causal variants (30% vs. 16%) and directional gene effects (43% vs. 23%) as the original AlphaGenome fine-tuned on single-cell data.
- The model was trained on more than 107 million individual cell profiles across 68 tissues.
- It predicts gene expression, histone modifications, chromatin accessibility, and DNA methylation.
- AlphaGenome 2 is available today for non-commercial research via the AlphaGenome API and in private preview for commercial research.
Conditional single-cell modeling lets researchers pinpoint variant effects in disease-relevant cell types without retraining the model for each new dataset.
HOPE: Higher-Order Pruning of Experts in Mixture-of-Experts Language Models
HOPE is a second-order expert pruning method that builds a pairwise expert interaction matrix F on a calibration set and solves a quadratic program to choose the prune-set, rather than scoring each expert independently as first-order methods do.
- Qwen3.5-122B has 256 experts in each of its 48 MoE layers, and about 95% of its ~250 GB parameter memory is taken up by expert MLPs, while only 8 experts per layer are used for any given token.
- At 50% pruning, HOPE achieves an average rank of 1.58 versus 2.42 for the next-best method, REAP.
- On Qwen3.5-35B-A3B at 50% pruning, HOPE scored 42.3 on SWE-Bench Pro, matching the unpruned model's 41.8, while REAP dropped to 36.2.
- The diagonal of the F-matrix recovers REAP's expert rankings almost perfectly (Spearman ρ = 0.988), so HOPE's gains over REAP come from the off-diagonal interaction terms.
For teams deploying large MoEs under memory constraints, HOPE shows that modeling pairwise expert cooperation preserves agentic coding ability at aggressive pruning rates, where independent expert scoring degrades sharply.
Token Pooling-Aware Training for Multi-Vector Embedding Models
A Hugging Face blog post explains pooling-aware training in Sentence Transformers, which applies a retrieval loss to both original and compressed document embeddings so the encoder preserves information that matters after token vectors are merged.
- With pool_factor=4, a normally trained model retains 94.6% of its unpooled NanoBEIR nDCG@10 with a 4x reduction in document vectors.
- Pooling-aware training raises that retention to 97.8% at the same storage budget.
- The loss is L = 1/2 L_unpooled + 1/2 L_pooled, with both objectives updating the same encoder.
- HierarchicalTokenPooling uses Ward hierarchical clustering and a straight-through estimator during training to pass gradients from pooled vectors back to their member tokens.
For multi-vector retrieval systems, pooling-aware training lets builders recover compression-related retrieval quality through training-objective alignment without increasing vector storage.
Train on prefill, monitor at decode: Using sparse autoencoders for topic monitoring
The paper introduces TopicScope, a framework for training causal topic classifiers on SAE features that assign a topic to every token so the label can change partway through a response. Across Gemma-3 1B and 4B, monitors trained on features extracted from prefill hidden states generalised to decoding.
- Three detector architectures were trained on synthetic multi-topic conversations covering seven topics.
- Across Gemma-3 1B and 4B, monitors trained on features extracted from prefill hidden states generalised to decoding.
- The classifiers were robust to perturbations, losing at most 5 macro-F1 points when off-topic asides were inserted, topics were removed or sections were reordered.
- Ablations on the contribution of the SAE found that classifiers reaching baseline performance without it were several times larger.
This shows compact topic monitors can be trained inexpensively at prefill and deployed during generation, and points to where SAEs are useful beyond safety classification.