Naresh Kumar Devulapally

VIBES

Hallucination Mitigation in Diffusion LLMs and VLMs using Validator guided Bi-Level Steering

Naresh Kumar Devulapally

Target: AAAI 2027

Summer 2026

VIBES

Naresh Kumar Devulapally

About VIBES

VIBES

Summer 2026

EMNLP Reviews

Naresh Kumar Devulapally

VIBES

Summer 2026

Reviewer - mr6y

in eq.8, the quality score is used additively to the score and not used 'only to break ties' as stated in lines 238

Answer: A better way to frame this answer will be to state that the selection of final candidate is dual objective where hallucination and quality are preserved with weights - check the code for weighted selection and include scalar weights with more weightage on hallucination and then quality.

EMNLP Reviews

Naresh Kumar Devulapally

VIBES

Summer 2026

Reviewer - mr6y

The update equations in algorithm 1 do not contain the terms Bgen or Bval, and therefore it is unclear how the constraints expressed in eq.2 are satisfied.

"The budgets are enforced by checks, not by a penalty term: the validator only fires when validator budget remains, and once either budget is hit the controller collapses to the cheapest path, so the totals in Eq. 2 can never be exceeded." - update algo 1: with b_gen, b_val values

EMNLP Reviews

Naresh Kumar Devulapally

VIBES

Summer 2026

Reviewer - mr6y

What are U and D, and the two lambda weights?

Write clearly the equations for each term in the updated version.

EMNLP Reviews

Naresh Kumar Devulapally

VIBES

Summer 2026

Reviewer - mr6y

in eq.6, \alpha_t at the right hand side of the equation appears different than the term in the rhs of the equation within (1+ \alpha_t)... Also the rhs \alpha, is never defined. In the algorithm 1, it appears to be the inspection budget and is initialized to 0. It is unclear how the inspection budget can be initialized to zero and then be increased at every iteration."

B_val` is the fixed budget (a constant); `n_val` is the running number of calls used (starts at 0, increases only on a call, and can never exceed `B_val`)."* Also make the two `α`'s the same font, or rename one

EMNLP Reviews

Naresh Kumar Devulapally

VIBES

Summer 2026

Reviewer - mr6y

in eq.6, \alpha_t at the right hand side of the equation appears different than the term in the rhs of the equation within (1+ \alpha_t)... Also the rhs \alpha, is never defined. In the algorithm 1, it appears to be the inspection budget and is initialized to 0. It is unclear how the inspection budget can be initialized to zero and then be increased at every iteration."

B_val` is the fixed budget (a constant); `n_val` is the running number of calls used (starts at 0, increases only on a call, and can never exceed `B_val`)."* Also make the two `α`'s the same font, or rename one

EMNLP Reviews

Naresh Kumar Devulapally

VIBES

Summer 2026

Reviewer - mr6y

in eq.6, \alpha_t at the right hand side of the equation appears different than the term in the rhs of the equation within (1+ \alpha_t)... Also the rhs \alpha, is never defined. In the algorithm 1, it appears to be the inspection budget and is initialized to 0. It is unclear how the inspection budget can be initialized to zero and then be increased at every iteration."

B_val` is the fixed budget (a constant); `n_val` is the running number of calls used (starts at 0, increases only on a call, and can never exceed `B_val`)."* Also make the two `α`'s the same font, or rename one