Neuromod-LLM: Treating Language Models Like Brains on Drugs

Hook

What if you could give an LLM the computational equivalent of caffeine, LSD, or a sedative—all without retraining, just by manipulating its attention patterns and memory cache during inference?

Context

Steering large language models toward desired behaviors typically requires expensive fine-tuning or brittle prompt engineering. You either commit GPU-hours to RLHF and maintain multiple checkpoints for different use cases, or you craft increasingly elaborate system prompts that models ignore half the time. Inference-time interventions—editing activations, scaling attention, pruning context—offer a third path: dynamic, reversible behavior modification without gradient descent.

Neuromod-LLM takes this concept and wraps it in neuroscience drag. The repository frames computational steering techniques as 'psychoactive substance analogues,' mapping biological neuromodulators (serotonin, dopamine, norepinephrine) to mathematical operations on transformer internals. A 'psychedelic pack' increases attention temperature to produce higher-entropy, more tangential outputs. A 'stimulant pack' reduces KV cache decay to enhance working memory and focus. The biological metaphor is scientifically dubious—there's no mechanistic homology between serotonin receptor binding and softmax temperature—but it organizes a useful taxonomy of steering primitives that practitioners might otherwise discover through trial and error. The real innovation isn't the neuroscience cosplay; it's the experimental rigor borrowed from clinical pharmacology, including double-blind evaluation protocols that could shame the field into better practices.

Technical Insight

Neuromod-LLM implements steering through a hook-based interception layer that wraps HuggingFace transformers. During forward passes, the framework applies 'neuromodulation packs'—bundles of parameterized operations targeting three architectural intervention points: residual stream injection (adding learned direction vectors to hidden states), attention pattern scaling (temperature-like adjustments to attention head softmax distributions), and KV cache decay (selective forgetting via geometric decay of cached key-value pairs).

The core abstraction is surprisingly clean. Here's how you'd apply a 'psychedelic' intervention that increases attention entropy:

from neuromod_llm import NeuromodulatedModel, AttentionTemperatureModulator

# Wrap any HuggingFace transformer
model = NeuromodulatedModel.from_pretrained("meta-llama/Llama-2-7b-hf")

# Define intervention: scale attention softmax by temperature factor
psychedelic_pack = {
    "attention_modulator": AttentionTemperatureModulator(
        temperature=2.5,  # Higher = more entropy, flatter attention
        target_layers=[12, 13, 14],  # Apply to middle-late layers
        fade_schedule="linear"  # Gradually increase through generation
    ),
    "residual_injection": None,  # Could add steering vectors here
    "kv_decay": None
}

# Apply pack and generate
model.apply_pack(psychedelic_pack)
output = model.generate(
    "Explain quantum entanglement",
    max_length=256
)

The attention temperature scaling is particularly clever. By multiplying attention logits by a temperature factor before softmax, the framework makes attention distributions flatter (higher temperature) or sharper (lower temperature). At temperature 2.5, the model attends more uniformly across context, producing outputs that human raters reliably label as 'tangential' and 'creative.' At temperature 0.5, attention becomes hyper-focused, yielding terse, repetitive text. This is underexplored as a steering primitive—most practitioners only adjust sampling temperature, not attention temperature.

The KV cache decay mechanism demonstrates a different failure mode by design. For 'sedative' or 'memory impairment' effects, the framework applies geometric decay to cached key-value pairs:

sedative_pack = {
    "kv_decay": KVCacheDecayModulator(
        decay_rate=0.92,  # Retain 92% per step, compounds over context
        start_position=32,  # Begin forgetting after 32 tokens
        layer_specificity=[0, 1, 2]  # Early layers = shorter-term memory
    )
}

This forces the model to 'forget' earlier context by exponentially reducing the magnitude of cached activations. The result is predictable degradation: repetition, incoherence, and failure to maintain long-range dependencies. It's a computationally cheap way to induce controlled failure—useful for studying how models rely on context, or for simulating constrained-memory scenarios.

The experimental design is where Neuromod-LLM diverges from typical ML research. Models receive randomized intervention codes, human raters score outputs without knowing which pack was applied, and statistical analysis uses mixed-effects models with false discovery rate correction. The custom psychometric instruments (PDQ-S for 'psychedelic' signatures, ADQ-20 for 'stimulant' effects) are questionnaires scoring outputs on traits like verbosity, tangentiality, and focus. This borrows clinical trial methodology wholesale—the double-blind protocol and structured behavioral readouts are genuinely novel for steering research, even if the underlying math is just vector addition and scalar multiplication.

The modular architecture means you can compose interventions: combine residual stream injection (add a 'creativity' steering vector extracted via PCA on contrastive prompts) with attention temperature scaling (flatten distributions to increase entropy) and tune the parameters for your use case. The hook-based implementation keeps interventions reversible—you can swap packs mid-conversation without reloading the model, enabling dynamic behavior switching that's impractical with fine-tuning.

Gotcha

The biological metaphor is marketing theater. There's no mechanistic connection between dopamine signaling and attention scaling—the 'biomimetic alignment' is purely behavioral analogy based on subjective human ratings of output traits. If you expect neuroscience insights, you'll be disappointed; this is vibes-based taxonomy, not computational neuroscience.

More critically, evaluation is entirely subjective. The psychometric instruments measure whether humans detect behavioral differences, not whether interventions improve performance on benchmarks like MMLU, GSM8K, or HumanEval. We don't know if 'psychedelic mode' produces creative breakthroughs or just entertaining gibberish—the framework could systematically degrade model quality while generating outputs that raters label as 'tangential.' There's no grounding in objective capability metrics, no ablation studies showing which interventions preserve reasoning, and no comparison against baseline performance. The intervention parameters (which steering vectors, what temperature factors, how much cache decay) appear hand-tuned or heuristically chosen. The repository mentions 'pack optimization' but doesn't implement it or document the tuning process, meaning you can't replicate their results or adapt packs to new models without trial and error. Scalability is also questionable: this requires full model access and per-token hook execution, making it incompatible with API-based models and potentially expensive for production. The computational overhead of attention scaling and cache manipulation on every forward pass isn't benchmarked—you're flying blind on latency costs.

Verdict

Use if: You're researching inference-time steering and want a modular framework for experimenting with attention scaling, activation editing, and context manipulation—the hook-based architecture is cleaner than rolling your own interception layer. You need reversible, parameter-free behavior switching for deployment scenarios where maintaining multiple fine-tuned checkpoints is impractical. You appreciate the double-blind evaluation methodology and want to borrow those ideas for your own alignment research, even if you ignore the neuroscience framing. Skip if: You need production-grade tools with API compatibility, care about benchmark performance preservation, or expect 'biomimetic' claims to mean anything beyond behavioral metaphor. You want principled parameter optimization instead of heuristic tuning, or you need evidence that interventions improve rather than degrade model capabilities. For most use cases, adjusting sampling temperature or using structured prompts achieves similar behavioral changes without surgery on model internals—Neuromod-LLM is a solution in search of a problem unless you specifically need fine-grained per-head attention control and can stomach the scientific cosplay.