diff --git a/CITATION.cff b/CITATION.cff
new file mode 100644
index 0000000..f05909f
--- /dev/null
+++ b/CITATION.cff
@@ -0,0 +1,16 @@
+cff-version: 1.2.0
+message: "If you use DecisionFormer in your work, please cite it as below."
+title: "DecisionFormer: Fine-tune, load and serve one-pass decision models on Hugging Face transformers"
+type: software
+authors:
+  - name: "FrontiersMind AI"
+repository-code: "https://github.com/FrontiersMindAI/decisionformer"
+license: Apache-2.0
+version: "0.2.0"
+date-released: "2026-10-24"
+keywords:
+  - decision models
+  - classification
+  - calibration
+  - transformers
+  - fine-tuning
diff --git a/README.md b/README.md
index 8dc6887..1324e56 100644
--- a/README.md
+++ b/README.md
@@ -324,9 +324,10 @@ truncate and refuse what does not fit instead.
 
 Bidirectional mode uses transformers' own switches (`is_causal=False` on the text decoder only), so SDPA,
 flash and sliding-window kernels all respect it, and a vision tower keeps its own attention. Conformance tests
-check every family in `tests/test_attention.py`. Compare against a causal baseline before adopting it: in our
-MNLI tests, SmolLM2-360M made bidirectional without adaptation stayed near chance where causal reached 0.80
-(`docs/benchmarks.md`).
+check every family in `tests/test_attention.py`. Compare against a causal baseline before adopting it: a causally
+pretrained decoder made bidirectional loses most of what it knows. On deciformer-15k, Qwen3-0.6B reached 0.45 after
+50 steps with bidirectional attention against 0.67 with causal attention, and on MNLI SmolLM2-360M stayed near
+chance where causal reached 0.80 (`docs/benchmarks.md`).
 
 ### Multi-GPU (DDP)
 
@@ -400,3 +401,17 @@ image and video inputs, and more heads (`multi_label`, `regression`).
 Apache-2.0. The option-code readout, prompt layouts, calibration metrics and near-duplicate checks are adapted
 from [AutoJev](https://github.com/denis-pplx/autojev) (MIT; see `NOTICE`). The `d3` prompt and checkpoint
 contract follow Decision 3.0 ([`vllm-sr/d3`](https://huggingface.co/vllm-sr/d3)).
+
+## Citation
+
+If you use DecisionFormer in your work, please cite it:
+
+```bibtex
+@software{decisionformer2026,
+  title  = {{DecisionFormer}: Fine-tune, load and serve one-pass decision models on Hugging Face transformers},
+  author = {{FrontiersMind AI}},
+  year   = {2026},
+  url    = {https://github.com/FrontiersMindAI/decisionformer},
+  note   = {Version 0.2}
+}
+```
diff --git a/docs/benchmarks.md b/docs/benchmarks.md
index 330a84e..3cf2f40 100644
--- a/docs/benchmarks.md
+++ b/docs/benchmarks.md
@@ -35,6 +35,19 @@ bidirectional attention and a mean pooler (0.84 above). With bidirectional atten
 (the Decision 3.0 layout, the default since v0.2), a 40-step QLoRA check on deciformer-15k showed the same loss spike
 near step 20 (1.55 → 2.04 → 1.23) and 0.41 dev accuracy on 128 records, so no Qwen3.5 recipe ships yet.
 
+**Bidirectional attention on a smaller decoder.** Three 50-step LoRA runs of the `deciformer15k_qwen3` recipe
+(Qwen3-0.6B, deciformer-15k, 256 dev records), identical except for attention and pooling:
+
+| attention | pooler | train loss, step 10 → 50 | dev accuracy |
+|---|---|---|---|
+| causal | last token | 1.55 → 0.73 | **0.672** |
+| bidirectional | last token (the Decision 3.0 layout) | 2.89 → 1.27 | 0.449 |
+| bidirectional | mean | 2.63 → 1.25 | 0.441 |
+
+Switching a causally pretrained decoder to bidirectional attention starts it from a much worse point (gradient
+norms 3-5x larger with the last-token readout) and it learns far more slowly, on Qwen3 as on Qwen3.5 and SmolLM2
+(v0.1, below). Bidirectional decoders need a long adaptation stage before they compete; keep `causal` otherwise.
+
 **Parity with the Decision 3.0 reference runtime.** `AutoDecisionModel.from_pretrained("vllm-sr/d3-edge")` against
 the checkpoint's own `trust_remote_code` runtime, BF16, four requests (choice, noul and score questions, an empty
 state, a JSON state, a 32-option choice): one question per forward pass, the probabilities are identical (maximum