attiq/blind-spots-qwen35-2b-base

Blind Spots of Qwen3.5-2B-Base

15 diverse failure cases of Qwen/Qwen3.5-2B-Base, a pretrained base language model released March 2, 2026.

Model

Model	Qwen/Qwen3.5-2B-Base
Type	Causal LM with vision encoder — pretrained base only (no SFT/RLHF/DPO)
Params	2B (1.88B measured)
Arch	Hybrid: 75% Gated DeltaNet (linear attention) + 25% Gated Attention
Context	262,144 tokens natively
License	Apache 2.0

Loading Code

Tested on Google Colab free T4 GPU. Model is ~4.5 GB in float16.

pip install "git+https://github.com/huggingface/transformers.git@main"
pip install accelerate torch datasets huggingface_hub pandas

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("Qwen/Qwen3.5-2B-Base", trust_remote_code=True)
mdl = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3.5-2B-Base",
    torch_dtype=torch.float16, device_map="auto", trust_remote_code=True,
)

def complete(prompt, max_tokens=150, temp=0.3):
    ids = tok(prompt, return_tensors="pt").to(mdl.device)
    with torch.no_grad():
        out = mdl.generate(
            **ids, max_new_tokens=max_tokens, temperature=temp,
            top_p=0.9, do_sample=True, pad_token_id=tok.eos_token_id,
        )
    return tok.decode(out[0][ids["input_ids"].shape[1]:], skip_special_tokens=True).strip()

All prompts use raw text completion (no chat template) since this is a base model.

Transformers must be installed from source — Qwen3.5's Gated DeltaNet architecture is not in any released version.

Schema

Column	Description
id	1–15
category	Failure category (e.g. arithmetic, commonsense, word_manipulation)
input_prompt	Exact prompt given to the model
expected_output	Correct answer
model_output	What the model actually produced (truncated for readability)
blind_spot_type	Description of the failure pattern

The 15 Blind Spots

1 — Self-contradictory arithmetic

"1000 − 37% = ?" → Says 621. Its own working correctly computes 370, but it subtracts wrong.

2 — Invalid syllogism

"All roses are flowers. Some flowers fade. Therefore…" → Affirms the consequent: "some roses fade quickly."

3 — Degenerate repetition loop

"iPad released ___ after the iPhone" → Loops "in 2010. The first iPhone was released in 2007." endlessly instead of computing the 3-year difference.

4 — Lateral thinking failure

"Pushed car to hotel, lost fortune" → Fails to recognize the Monopoly riddle. Enters <think> mode and speculates about metaphors.

5 — Think-tag leak / no final answer

"Facing north, turn right 90°, turn 180°" → Enters <think>, reasons correctly through compass math, but never outputs "west".

6 — Self-contradictory physics

"Bowling ball vs feather in vacuum" → Says "the bowling ball", then explains in detail that both fall at the same rate.

7 — Format hallucination

"Schadenfreude means…" → Invents a multiple-choice quiz nobody asked for, leaks <think> tags, never gives the translation.

8 — Instruction non-compliance

"List exactly 3 fruits" → Outputs "2. 3." then enters <think> mode reasoning about fruit but never lists any.

9 — Ignored constraint

"Snail climbs 3m/day, slides 2m/night, reach 10m?" → Says 4 days, completely ignoring the nightly slide-back (correct: 8).

10 — Counterfactual override

"Sun rises in west, I should face…" → Says "the east", overriding the stated premise with real-world knowledge, then loops.

11 — Character reversal error

"Reverse 'parts'" → Says "traps" (anagram) instead of "strap" (true reversal). Tokenization prevents char-level ops.

12 — Vowel removal error

"Remove vowels from 'education'" → Says "edcatn", keeping e and a. Same tokenization root cause.

13 — Confident hallucination

"Snake legs?" → Says 4, then writes "snakes are legless reptiles", then invents hidden legs beneath scales.

14 — Trick question failure

"3 apples, take 2, how many do YOU have?" → Says 1 (3−2 remaining), missing that YOU now possess 2. Then loops.

15 — Self-contradictory word problem

"3−1−1+5 = ?" → Says 8, then step-by-step derives 6, then invents a "phantom apple" to explain the discrepancy.

Failure Patterns

Pattern	IDs	Root Cause
Self-contradiction (answer ≠ explanation)	1, 6, 13, 15	Generates answer token first, then reasons — cannot backtrack
<think> tag leakage	4, 5, 7, 8	Pretrained with control tokens but never trained to use them
Repetition loops	3, 10, 14	Gated DeltaNet linear attention may lock into fixed hidden states
Tokenization blindness	11, 12	Subword tokenization prevents character-level manipulation
Constraint ignoring	9, 10	Skips stated constraints (slide-back, counterfactual premise)

Recommended Fine-Tuning

Datasets

Blind Spot	Training Data
Arithmetic	GSM8K, MATH
Logical reasoning	LogiQA, FOLIO
Commonsense	CommonsenseQA, PIQA
Spatial reasoning	SpartQA, custom rotation tasks
Character-level ops	Synthetic reversal, counting, and vowel-filtering tasks
Instruction following	Tulu-3, OpenAssistant

Assembly

1. Curate from existing benchmarks (most are free on HuggingFace)
2. Synthetic generation via stronger models for chain-of-thought examples
3. Adversarial filtering — run base model, keep only failures
4. Human curation for commonsense and trick questions

Size estimates

At 2B params, quality > quantity:

• Targeted fix (5 patterns): 10K–50K examples with chain-of-thought
• Broad improvement: 50K–100K multi-task
• Comprehensive: 200K–500K (cf. Tulu-3, SmolLM3)

Qwen3.5-2B-Base supports efficient LoRA via pretrained control tokens (<|im_start|>, <|im_end|>) without finetuning embeddings.

License

Apache 2.0