HybridAgent 8B beats GPT-5.6 Luna on GAIA by reasoning in hidden states, not tokens

Researchers from Stanford, UIUC and Together AI have released HybridAgent, a family of open-weights agents that can do part of their reasoning without writing any text. According to a dev.to write-up of the work, the paper — posted to arXiv on October 13, 2026 — reports that the 8-billion-parameter model scores 43.7 on the GAIA agent benchmark, ahead of GPT-5.6 Luna at 37.9 and Kimi K3 at 41.7. The same results table puts it at just 14.4 on Humanity's Last Exam, where Luna scores 29.0.

The release is fully open: code, weights, and the SFT and RL datasets are published under MIT. Two checkpoints, HybridAgent-8B-Agentic and HybridAgent-14B-Agentic, are on Hugging Face, both built on Qwen3 base models. The agents call a Python interpreter and a web search tool, with search requiring a Tavily API key. The paper reached the Hacker News front page on October 18.

How the latent mode works

A conventional agent reasons in text, paying for every token of its thinking. HybridAgent can leave text behind. When the model emits a special latent token, the following steps skip decoding: the last-layer hidden state passes through a small residual adapter the authors call HybridLink (roughly 10 million parameters on a 4B model) and is fed back in as the next input embedding. The latent span ends when that output lands closest to the closing token's embedding, and nothing fixes the span length in advance — the model decides as it goes. Between turns, the agent keeps the KV cache of everything so far, including latent steps and tool results, rather than re-encoding the trajectory as tokens. A reinforcement-learning stage, HybridRL, teaches it when to think quietly and when to write.

The paper's abstract claims, across three model scales and eleven benchmarks, an average 10.4% improvement, 35.5% to 54.7% fewer tokens, 1.8x to 2.6x faster inference and 2.7x to 3.7x lower cost.

A scorecard with sharp edges

The averages hide uneven rows. The small agents dominate web navigation: on WebWalker, the 8B and 14B models score 42.5 and 49.5 against 11.0 for Luna and 10.0 for Kimi K3. But on Humanity's Last Exam, a hard knowledge-and-reasoning test, the frontier models score roughly double, and on DeepSearchQA Luna still leads with 23.3 accuracy.

One further comparison complicates the headline. ARPO, a text-only 8B agent trained with multi-turn tool-use RL and no latent mode, already scores 38.8 on GAIA and 32.0 on WebWalker — above Luna on both. Beating Luna on GAIA is therefore not unique to latent reasoning; what the latent mode adds is margin, lifting GAIA from 38.8 to 43.7 and WebWalker from 32.0 to 42.5.

Fine print on the cost claims

The cost figures price compute, not an API bill. The paper's estimates assume $2.50 per H100-hour at 10% of peak BF16 throughput, charge each latent step as a full forward pass, and exclude prompt prefill, context-dependent attention and external tool execution. On those terms the 8B model's cost drops from $439 to $121 per million questions, with search API charges and prefill on top.

Reasoning you cannot read

Latent spans produce no text, which strains observability setups built on reading reasoning traces. A Hacker News commenter raised exactly this, asking whether the approach weakens monitorability and safety. The authors offer a partial window: each latent step can be mapped to its nearest token embedding to produce what they call pseudo-text, and the repo ships an inspector that renders these spans as an HTML report. But a nearest-token match shows what the hidden state points toward, not what the step actually computed, and the paper notes that small errors can compound across a long latent span and steadily degrade the coherence of the latent trace.

Trying it today

Practical barriers remain. The model card warns that loading with transformers alone yields a text-only model; latent mode and tools require the HybridAgent code. No inference provider hosts it yet, and the authors say verl and vLLM do not directly support multi-turn hybrid generation, so they extended their own stack. Training HybridRL is recommended on 32 GPUs. One Hacker News commenter who ran the 14B model outside the official engine reported it looping on repeated paragraphs until its turn budget ran out, suspecting a llama.cpp incompatibility — which, given the model card's warning, suggests such runs never exercise the hybrid part at all.

Why it matters

The result is a useful corrective to the assumption that agentic benchmarks require frontier-scale models: an 8B open-weights agent, with fully public training data, beats a leading proprietary model on GAIA and outscores it several times over on web navigation. But the details matter as much as the headline. The wins concentrate in search-heavy tasks, a strong text-only baseline of the same size already beats the same frontier model, and the efficiency gains trade away something most agent deployments rely on — reasoning a human can read. For workloads where you check the answer rather than the path, that trade may be worth it. For anything that must be audited, text traces still carry real weight.